Recommendation Systems for e-Shopping: Review of Techniques for Retail and Sustainable Marketing

Abstract

In recent years, the interest in recommendation systems (RSs) has dramatically increased, as they have become main components of all online stores. The aims of an RS can be multifaceted, related not only to the increase in sales or the convenience of the customer, but may include the promotion of alternative environmentally friendly products or to strengthen policies and campaigns. In addition to accurate suggestions, important aspects of contemporary RSs are therefore to align with the particular marketing goals of the e-shop and with the stances of the targeted audience, ensuring user acceptance, satisfaction, high impact, and achieving sustained usage by customers. The current review focuses on RS related to retail shopping, highlighting recent research efforts towards enhanced e-shops and more efficient sustainable digital marketing and personalized promotion. The reported research was categorized by main approach, key methods, and specialized e-commerce problems addressed, while technological aspects were linked with marketing aspects. The increasing number of papers in the field showed that it has become particularly popular, following the explosive growth in e-commerce and mobile shopping. The problems addressed have expanded beyond the performance of the core algorithms to the business aspects of recommendation, considering user acceptance and impact maximization techniques. Technologies have also shifted from the improvement of classic filtering techniques to complex deep learning architectures, in order to deal with issues such as contextualization, sequence-based methods, and automatic feature extraction from unstructured data. The upcoming goals seem to be even more intelligent recommendations that more precisely adapt not only to users’ explicit needs and hidden desires but also to their personality and sensitivity for more sustainable choices.

1. Introduction

Recommendation systems (RSs) are nowadays widely adopted in all fields of e-commerce, following the explosive growth in the variety of offered products, their providers, and the information sources about the available choices. New e-business services such as product search and comparison websites, auctions, and review blogs have overwhelmed consumers, creating a pressing need for recommendations. RSs are valuable means for increasing the visibility of consumption trends and niche markets, and for guiding customers, especially those with particular preferences, such as sustainability-minded persons. RSs are able to point a potential customer towards relevant and well-suited choices, using knowledge about customers, either expressed explicitly as ratings and comments or inferred from their actions. The importance of RS in e-commerce is even clearer from the viewpoint of businesses, since attracting customers, building loyalty, and increasing engagement and satisfaction become essential aspects of a business strategy. An RS can greatly contribute to such business goals by creating a value-added relationship between the site and the customer. The depth of insight gained into a customer’s preferences and interests can thus become a strong competitive advantage. In their recent study on economic sustainability and the effect of e-commerce, Kennedyd et al. [1] highlighted the importance of building shopping site commitment and customer loyalty, rather than intention for single purchases. In addition to other influencing factors, it was found that building a trustworthy RS and tailoring the e-shop to local cultural tastes largely contribute to the achievement of repeating purchasing and to the establishment of sustained e-shopping platforms.

Recommendation systems emerged as an independent research area in the mid-1990s [2,3,4]. For any business-to-customer (B2C) company, high-quality recommendations have always been important [5], since they can effectively enhance the customer’s trust, satisfaction, and loyalty [6]. In recent years, the interest in RSs has dramatically increased, as they have become main components of online stores, and they also play an important role in the marketing actions of physical stores. One of the most successful cases is Amazon.com, which started from book recommendations and expanded to the employment of a sophisticated RS to personalize its online store for each customer [7], achieving a huge increase in their sales. In their article titled “Two decades of recommender systems at Amazon.com”, Smith and Linden [8] referred to the main principles of their approach and the value that it still holds. The basis of their algorithm was to maintain a precalculated table of “related” items for every product in the shop’s catalogue. Customers who express interest in certain items then receive recommendations from the readily available list of related items. The selection of the most suitable items, as explained by the authors, can only be successful if important factors are considered, such as time aspects and the nature of the market field.

The state of the art in the wide field of RS technologies has been captured by a large number of reviews found in the literature. Many of the existing reviews focus on e-commerce, emphasizing either on the business aspects of the application or on the algorithmic perspectives. A gap that has been identified was to link technological aspects with marketing aspects, considering that the goals of an RS can be related not only to the increase in sales or the convenience of the customer, but may include the promotion of alternative environmentally friendly products or to strengthen policies and campaigns. In this paper, we focused on RS related to retail shopping, considering developments towards enhanced e-shops and more efficient digital marketing actions within loyalty programs, store mobile apps, or personalized promotion. Our goal was to capture the most promising approaches and latest trends in developing high-performance solutions towards business-oriented achievements. The scope did not include content recommendations (films, music, articles, etc.), e-tourism, or the broader e-business sector. The intention of the current review was to depict notable efforts in improving the performance of RS in the context of retail business, rather than to map the entire landscape of RS technologies.

To this end, the business goals, sustainable marketing problems, and major challenges of RS in retail shopping were captured and were associated with desirable method characteristics and suitable high-performance solutions reported in the literature.

The contribution of the current review was to narrow the field of e-commerce recommendation applications and to present a well-focused landscape of the particular section of retail shopping, further highlighting research relevant to sustainable marketing (e.g., promotion of environmentally friendly items). In addition, business goals were associated with relevant evaluation metrics and performance aspects. The reviewed papers were categorized in three dimensions: general approach, main algorithm, and problem addressed. In each dimension, we referred to important method characteristics towards business-oriented achievements.

In the next section, the review methodology is described, and in Section 3, prior reviews on RS in e-commerce are presented. In Section 4, we summarize the major goals, challenges, and business-oriented performance aspects. Section 5 outlines typical RS applications, organizing the reported research work according to its main approach or modeling method, while Section 6 covers more complex recommendation problems with a particular interest in retail. Finally, in Section 7, the current research landscape is summarized, and the remaining challenges and future trends are discussed.

2. Methodology

The methodology of the current review included three steps (Figure 1):

(1)

A wide initial search was performed in Google Scholar and Scopus, using the keywords “recommendation systems in retail”, “recommendation systems in e-commerce”, ”recommendation systems for e-shops”, and ”recommendation systems in sustainable marketing”. The publication year and the number of citations were recorded, and the results were processed by the authors to remove duplicates, papers that were not relevant, and those with minimal impact (i.e., with fewer than three citations). It is noted that additional searches in other major databases did not bring out any additional findings.

(2)

The papers returned by the search engines—both review and research papers—were studied in order to extract the main applications, popular problems in retail RS, and challenges addressed, as well as the typical and emerging algorithmic approaches employed in this area. The references of these papers were also tracked to find additional articles of interest. The findings were used to define the focal points of the relevant research, which was mapped on three dimensions: (a) the main approach (e.g., collaborative filtering, content-based, knowledge-based), (b) the key methods (e.g., neural networks, rule-based, matrix factorization), and (c) the particular marketing problem addressed (e.g., user acceptance, contextualization, explainability). In parallel, a set of additional keywords was identified, which reflected more specific and detailed terms in the area of interest.

(3)

An extended search using the specific keywords of the previous step was then performed to dig out any missed-out papers addressing specialized problems. All the identified papers were filtered to keep those that were indeed contributing relevant information to the area of interest and demonstrated considerable impact. The selection was based on the citations received in relation to the years since publication, how recent they are, and how closely targeted they are to the narrowed field of our review. The research found was classified according to the dimensions and categories of step 2. Finally, the filtered and classified papers were reviewed by the authors in more detail, in order to compile a comprehensive report on the notable efforts in RS for retail shopping.

The initial search resulted in 23 review papers and 397 research papers. Through the search with additional keywords and by tracking the references of these papers, another 43 relevant papers were found. Papers that were focused on applications other than retail shopping were filtered out; however, we kept selected articles with notable information on relevant RS methods, even if they were not focused on the particular application field.

3. Prior Reviews of RS in e-Commerce

3.1. Reviews of Recommendation Applications

Before getting into the review of recommendation applications, it is worth clarifying the term recommendation system in the e-commerce context and differentiating it from other similar terms. Li and Karahanna [9] started their article by specifying that RS is a system that collects a consumer’s preferences using data on purchase history, ratings, and profile in order to recommend tailored products or services predicted to best suit these preferences. They differentiate recommendation from interactive decision aid systems, which are defined as a more general category, where consumers are directly asked for their preferences. Personalization and customization have a similar but wider meaning, covering technologies from tailor e-commerce interactions to consumers’ preferences. In the same paper, the authors focus on 41 empirical studies until 2013, and organize them according to the three-stage model originally proposed by Adomavicius and Tuzhilin [10]: (1) collecting consumer data and building consumer profiles, (2) matching products or services to consumer profiles and presenting those recommendations, and (3) understanding and measuring the impacts of these recommendations and adjusting personalization strategies based on this feedback.

In a recent review targeting RS in e-commerce, Alamdari et al. [11] selected 33 papers from a much larger number of papers in this field from the period 2008–2019. They briefly discussed the RS evaluation metrics, which are particularly relevant in e-commerce. After categorizing the papers according to their basic approach, they provided a comparative overview of the main advantages, disadvantages, and unsolved problems in each category. Fayyaz et al. [12] presented the state of the art in RS in 2020, and although they encompassed a wide range of applications, they made special reference to business aspects and applications in e-commerce. The algorithmic approaches were briefly presented together with critical comments on each one’s advantages and limitations, while the main challenges in RS were outlined together with important efforts for their solutions.

In an older but very interesting survey on the latest developments of e-commerce RS, Karimova [13] gave emphasis to the perspective of e-vendors. Papers were presented that aim at specialized e-shop problems, such as increasing customer loyalty, maximizing revenue by balancing the probability of purchasing a product with the corresponding revenue, considering specialized RS for small and medium e-shops, and addressing long-tail markets. Xiao and Benbasat [14] reviewed empirical studies on the use of recommendation agents from the period 2007–2012. The issues considered were the perceptions of users and the way they affect the acceptance and use of the recommendation systems. Taking up on these findings, the authors refined a previously proposed model on the use of product recommendation agents for retail stores.

3.2. Reviews of Recommendation Technologies

In an extensive recent review by Singh et al. [15], more than 1000 papers were identified from 2011 to 2017. Although not specific to e-commerce, this review paper was very useful in matching the most popular recommendation methods with major application domains, as well as with the major addressed problems. Their predictions regarding trends in modeling included an emphasis on reinforcement learning and extensions of recurrent neural networks to enhance the contextualization abilities of RS. A more specialized review was published by Almahmood and Tekerek [16], focusing on deep learning algorithms applied to online shopping RS. Comparing the reported evaluations for various relevant algorithms, they concluded that solutions based on deep neural networks, e.g., CNN and RNN (see Section 5), perform better in terms of typical recommendation problems, such as cold start and sparsity.

Zhang et al. [17] reviewed the applications of artificial intelligence in RS, identifying current research issues and indicating new research directions. In their search, they categorize and present research work in the field of AI that is related to the enhancement of recommendation techniques. They refer to techniques such as transfer and active learning, deep neural networks, natural language processing, and computer vision, and review how these are introduced in RS. As areas open for additional research, they mention the time awareness (to forget older data and focus on recent customer behavior), consideration of long-tail items (specialized items that are addressed to limited audience), privacy issues, and visualization, explainability, and interactivity issues of the RS interfaces.

In order to study the landscape of knowledge-based recommendation systems (KBRS), Bouraga et al. [18] proposed a classification framework that organizes the reported work based on the RS features in multiple dimensions. They classified KBRSs according to the recommendation problem and solution (i.e., nature and content of the knowledge base and whether the recommendation algorithm is quantitative or qualitative), user profile (based on implicit or explicit input and whether the user profile is known from historical data or learned through the current actions), and the degree of automation. Kim and Chen, in their scientometric review published in 2015 [19], applied a computational approach and an expanded research in order to capture emerging trends in RS research. They identified a large set of keywords and references; they considered the dimension of timing and employed the scientometric software CiteSpace [20] as a visualization tool. In this way, they were able to identify “citation bursts” and “keyword bursts” and, therefore, timelines of research themes and emerging trends. While, until 2008, personalization was by far the strongest keyword, in 2014, the strongest keyword bursts were social media and word of mouth.

Finally, for historical tracking purposes, it is interesting to mention the review of 240 articles from 2001 to 2010 by Park et al. [21], who classified them per application area and data mining technique, using a solid classification process. Although not restricted to e-commerce, their classification by application field showed that 20% of the papers addressed the area of e-commerce, which was mostly based on association rules, k-NN, clustering, and various heuristic methods. The authors noticed that social network analysis was underdeveloped and predicted that it was a promising field for the future, something that has nowadays been confirmed.

4. Goals and Major Challenges of Recommendation Systems in e-Commerce

The success of an RS is multifaceted, being related to the value it offers either to the commercial company that runs the system, to the customer, or to both. Its potential goals include acceptance, satisfaction, and sustained usage by customers, as well as effectiveness in increasing sales or engaging the customer to a commercial site. In this section, we review the business goals at which an RS may be aimed, in order to be of high value to the commercial platform in which it is introduced. These goals are then translated into the metrics that a recommendation algorithm should optimize, according to its usage context.

4.1. Business-Oriented Performance Aspects

An important issue that has been established in e-commerce is that recommendations should not be simple reflections of customers’ purchase history but should offer alternatives, encouraging them to try new products outside the circle of their regular buys. In addition to covering the customer’s regular needs, spontaneous buys in entirely new areas should also be stimulated. In this way, the users’ goal to receive assistance during their shopping is served together with the e-shop’s goals for up-selling and triggering pleasant feelings of happy discoveries [22]. The quality of the recommendation is therefore not entirely dependent on the accuracy of the prediction algorithm but on a multitude of factors related to consumer behavior and marketing aspects [23]. The focus of the RS is often dependent on the particular commercial sector; e.g., in eGrocery, the recommendations may be focused on relevance to the user’s needs, while in clothing to the discovery of fresh choices. The user’s expectations may also depend on his/her personality or the context of a particular visit to an e-shop; e.g., a potential buyer of shoes may value familiarity to a particular style/brand and wish to stay within a comfort zone, another may prefer diversity but only within products relevant to his/her needs, while a third one may wish to explore choices that are not even relevant to the user’s purchase history.

Another point to be considered is the gap between the intention of the customer and his/her actual actions. This problem is particularly relevant to sustainable marketing, where a customer may be willing to prefer an environmentally friendly product but finally purchases a conventional one. The reason may be the difficulty in finding appealing choices but also the lack of convincing information about the sustainability characteristics of candidate products [24].

The most well-established concepts to capture RS qualities other than accuracy are diversity, novelty, and serendipity [11]. They express the need that recommendations should not be obvious or static repetitions of a customer’s history and should not suffer from the overfitting problem, i.e., to recommend only the specific items highly associated with a user, without any chance of suggesting something new. Serendipity adds the need to create positive feelings. Adamopoulos and Tuzhilin [25] further introduced the concept of unexpectedness, differentiating it from novelty, serendipity, and diversity. They formulated the concept theoretically and proposed performance metrics to evaluate it. They developed a method for increasing user satisfaction by providing recommendations that are both unexpected and accurate regarding relevance to the user’s interests.

Considerable work has been reported on increasing diversity and tuning it to the users’ needs since as early as 2001. Lawrence et al. [26] proposed an RS application for supermarket customers aiming at providing them with “new ideas” beyond their regular purchase patterns. Eskandanian et al. [27] aimed at matching the RS diversification to each user’s preference, using a profile clustering approach. A wide study on the problem of introducing diversity has been elaborated by Kunaver and Požrl [28], where papers have been reviewed on the definition and evaluation of the concept of diversity, on the impact of the diversification process on the overall quality of an RS, and on the related algorithms. Jugovac et al. [29] addressed the problem of optimizing an RS based on multiple quality factors, including accuracy and diversity, as well as individual user tendencies. To this end, they presented a parametrizable multiobjective reranking method.

Additional success factors in the e-commerce domain are related to the customer’s expectations, perceptions, and acceptance of the entire process, rather than the recommendation list itself. Studies on the expectations of customers from personalized recommendations [30] identified and evaluated 14 customer expectations that have to do with the recommendation outcome itself, e.g., accuracy (to match their preferences) and discovery (to help them discover items they would not find); the process (e.g., to offer a convincing connection between customer’s preferences and the recommended products); and several others, including the customer’s and the marketer’s roles. Interesting findings that can be considered in the RS design were that product knowledge, customer knowledge, and sales motive are important expectations. This implies that rough suggestions that only match the interests of the company rather than the real needs of the customer should be avoided. It was also confirmed that customers emphasize the accuracy benefit when they have solid preferences but prefer the discovery benefit when they do not have clear preferences or when they believe that they have preferences that they wish the RS would help them discover. A personalized RS could therefore adapt to the preference development of individual customers.

Other studies have shown that successful personalization can significantly increase positive attitudes, such as satisfaction and enjoyment [31]; however, customers who are overloaded with information are likely to develop negative emotions [32]. Translating these findings into RS parameters, a developer may balance his algorithm in favor of high sensitivity at the expense of accuracy and at the same time limit the number of recommendations to fit the tolerance of the targeted audience. Additionally, Gai and Klesse [33] suggested that user-based vs. item-based framing plays a role in the effectiveness of recommendations. User-based framing emphasizes the similarity between customers (e.g., “people who like this also like…”), while item-based framing emphasizes similarities between products (e.g., “similar to this item”). In their experiments, they found that user-based framing generally achieves better results, irrespective of the accuracy of the algorithm itself. The overall effect of personalized recommendations may also be dependent on service delivery aspects, other than the performance of the core algorithm, such as the timing of delivery and the image of trust [34].

Finally, it is worth mentioning that within the success factors of an RS, there are several design aspects, which are totally independent from the recommendation engine, such as the aesthetics of the interface and the functionality offered to the e-shop customer. These aspects are outside the scope of the current review; however, it is noted that they attract significant research efforts, such as the work of Bortko et al. [35], who conducted eye-tracking experiments to optimize the user experience of the recommendation interface.

4.2. Evaluation Metrics for Recommendation Algorithms

The evaluation of an RS is not trivial and should be aligned with the business goals, while the selection of the criteria heavily affects the results. Evaluation includes the offline evaluation, which is performed in the lab by the algorithm developer using—preferably—real historical datasets, and online evaluation, which is performed in production, based on the reaction of real customers. In their review of evaluation metrics for RS, Silveira et al. [36] explored various evaluation concepts and proposed metrics or strategies to evaluate recommendations. In the following paragraphs, we outline the traditional and most recent metrics most suitable for e-shops, commenting on the suggested use of each one.

4.2.1. Accuracy Metrics

Evaluating Like/Dislike Predictions

Selecting items that match the preferences of a target user is the prediction whether a user likes or dislikes an item, which can be considered as a classification problem with two possible classes: “like” and “dislike”. Therefore, typical classification metrics are used to evaluate RS performance, such as accuracy, precision, recall, F-measure/F1, and ROC/precision–recall curve [37]. The simplest measure is Accuracy, defined as the percentage of successful recommendations over all the produced recommendations. It is of limited scope, since it does not provide information about the type of error that the algorithm is prone to. More informative metrics are Precision and Recall. Precision expresses the percentage of the cases classified as positive that are truly positive (i.e., how confident we are that a recommended item is indeed liked), and recall (or sensitivity) expresses the percentage of the cases correctly classified as positive over all the cases that are truly positive (i.e., how confident we are that we did not miss out liked items). In order to describe the total performance of the classifier in one measure, it is convenient to combine precision and recall in the criterion F-measure, which in its general form is defined as

$$F_{\beta }=\frac{Precision·Recall}{(1-\beta )·Precision+\beta ·Recall}$$

(1)

where $\beta$ balances the weight in favor of precision or recall. A very common choice is to equally balance the two metrics by choosing $\beta =1/2$. This measure is called F1-score and equals the harmonic mean of precision and recall:

$$F1=\frac{2·Precision·Recall}{Precision+Recall}$$

(2)

Depending on the marketing goals, the most representative measure can either be Precision, Recall, or the F-measure, that best balances them [37]. Emphasis on precision can be given when we want to recommend “some good items”, i.e., when we expect that there are a large number of matching items and our goal is to discover some of them or identify the best of them. In this case, the number of good items that we did not recommend is not a meaningful measure. Recall is more important when we want to recommend “all good items”, assuming that there are limited matching items that we do not want to miss. It is noted that the above metrics emphasize the correct or nondetection of the class “Like”, assuming that we only care about successfully choosing the items that the user will really like, rather than ruling out the ones that he would not like. The former is typically the case in e-shops; however, in certain applications, it is relevant to consider metrics that focus on the negative class (dislike). Such a measure is specificity, also known as true-negative rate or fallout, which shows at which percentage the RS correctly classifies items as “not matching”.

Evaluating the accuracy of rating predictions is typically performed using regression evaluation metrics, such as the “mean squared error (MSE)” or the “mean absolute error (MAE)” [37]. However, this case is not common in RS for retail shopping because, unlike application areas such as movies, the goal is rarely to predict a quantitative rating variable.

Evaluating Recommendation Lists

It is very common in e-commerce that recommendation algorithms offer ordered lists of recommended items, e.g., when the e-shop presents the five best offers that match the interests of the user. In such cases, the entire ordered list is evaluated, rather than individual items.

The Mean Reciprocal Rank (MRR) evaluates the average quality of the recommendation list over Q user queries. For a single query i, the reciprocal rank is defined as $R{R}_i=\frac{1}{ran{k}_i}$, where $ran{k}_i$ is the list position of the item actually chosen by the user. The highest value for the reciprocal rank is $R{R}_i=1$, when the user selects the first recommendation in the list, while $R{R}_i=0$ if the user chooses no item from the recommendation list. The mean reciprocal rank is the average over all recommendation lists:

$$MRR=\frac{1}{Q}\sum _{i=1}^{Q}R{R}_i$$

(3)

Precision and Recall@k can be used when the RS offers a list of the top-k recommendations, where k is the length of the list, which is fixed to match the site’s needs:

$$Precision@k=\frac{Recommendeditems@kthatarerelevant}{k}$$

(4)

$$Recall@k=\frac{Recommendeditems@kthatarerelevant}{Totalofrelevantitems}$$

(5)

The length of the recommendation list k is expected to affect a trade-off between precision and recall for an algorithm of a given quality. Increasing k is expected to favor recall because the RS has more chances to successfully hit more matching items. A smaller list length may result in higher precision because the RS will only recommend the few best matches with the highest probability of being correct. In order to have a clear view of an algorithm’s abilities, it is common to measure precision and recall at several different values for k. It is worth mentioning that in some RSs, the length of the recommendation list can be left flexible for the algorithm to decide per case so that the RS recommends as many items as it considers “good matches” based on a threshold to a user–item distance measure [38]. In such cases, an overall precision, recall, and/or F-measure can be evaluated at variable k.

Considering that the adjustment of certain algorithm parameters (including the length of the recommendation list) affects a trade-off between precision, recall, and false-positive rate, an expressive picture of the overall performance of the RS in the entire range of parameter settings can be obtained using the Precision–Recall Curve and the Receiver Operating Characteristic curve (ROC) [37]. The Precision–Recall Curve compares Precision (i.e., that the recommended items are indeed good matches) with Recall (i.e., with whether all the available good items are recommended). It is relevant when it is desirable to produce many successful recommendations, while some bad recommendations are not expected to have a negative impact. This is usually the case in recommendations provided during the browsing activity of e-shop customers. On the other hand, the ROC curve compares precision with fallout (i.e., with the recommendation of bad items), which is more representative when unsuccessful recommendations may be annoying, such as in push notifications and personalized promotion messages [12].

Finally, an important classification metric, which is used quite often lately, especially for comparing models, is Log-loss. In addition to the correctness of the predicted class, this metric considers the confidence with which the classifier made this prediction. Log-loss penalizes the incorrect classifications more heavily when they are predicted with high confidence. It is calculated as

$$Log-los{s}_i=-[y_{i}ln\left(p_i\right)+(1-y_i)ln(1-p_i)]$$

(6)

where i is the given observation, y is the true value, and p is the prediction probability. In order to compare models, the average Log-loss is calculated over a testing dataset. Lower Log-loss values indicate better performance.

4.2.2. Evaluation of Business Aspects of Algorithm Quality

Several metrics have been proposed for evaluating other than accuracy properties of the recommendation algorithms, with an impact on their success in e-commerce applications. The most representative ones are diversity, novelty, and serendipity.

Diversity can be evaluated through numerous proposed approaches, as presented by Kunaver and Požrl [28]. It has been measured as the average dissimilarity between all pairs of items in the recommendation list or using the Gini index as a measure of inequality in the item frequencies, or even by empirically surveying user perceptions. Efforts have also been reported to combine diversity with accuracy in a unified performance measure.

Novelty measures the degree in which the recommendations are unknown to the user, which is applicable when products known to the user or already bought are useless. It is noted that novelty is not useful in applications where a promotional action for a known (or even favorite) consumer product would be desirable. Novelty can be measured based on the distance of the recommended items from the target user’s history [25].

Serendipity is related to diversity and novelty but is differentiated by its focus on the feeling of pleasant surprise. Its measurement is usually performed in online evaluations with questionnaires by real users [39]. As an objective measure of serendipity that can be used for algorithm tuning, Ge et al. [40] proposed to measure unexpectedness by comparing the list to be evaluated with the items produced by a primitive model that tends to output fully predictable items, and then to combine unexpectedness with usefulness.

Other metrics that have been proposed are

Popularity evaluates whether the output of the RS is dominated by the few globally popular items. While large popularity is not necessarily undesirable, it may show the inability of the RS to provide high-value personalized recommendations.

Coverage measures the number of items that the RS was able to propose out of the total available items. Ge et al. [40] further defined the terms prediction coverage and catalogue coverage and proposed corresponding metrics. Small coverage indicates a smaller business value both for customers and for the shop, since a large number of available products or offers will never become visible through the RS.

4.2.3. Metrics for Online Evaluation of Business Value

The business metrics are used in online evaluation and measure the effectiveness of the RS in achieving the e-shop’s business goals, such as higher engagement of the customer, increased loyalty, up-selling, or cross-selling. The simplest metrics that can easily be measured are

Click-through rate (CTR). Assuming that a commercial site presents to its users a recommendation list N times, CTR is the proportion of the cases where the user clicks on a recommended item. This measure captures the ability of the RS to attract the attention of the user and to increase his interaction with the site. High CTR is a good indication that the recommended items were found interesting and the RS was of some value to both the e-shop and the user.

Conversion rate (CR). It measures the tangible impact of the RS on sales as the proportion of the customers who responded to the recommendation by adding the item to the cart and/or proceeded in buying it.

A more representative but difficult-to-measure metric is the effect on customer lifetime value (CLV). It expresses the degree in which users find useful products, they are satisfied by the overall purchase experience, they proceed in increased buys, and they are expected to return in the future.

5. Recommendation Types and Methods in e-Retail

5.1. Typical Applications of RS

The applications addressed in e-retail, as found in the current market landscape, can be categorized as follows:

• An e-shop recommending products to its visitors. Important characteristics of the problem are: (a) Whether the recommendations are produced before or after a sale or an item selection action by the user. A recommendation that precedes the user’s action is based on the general needs of the users, while the one that follows a user’s selection aims at matching the current user’s action. (b) Whether the recommendations are addressed to known or unknown users. When the user is unknown, recommendations can only be based on general knowledge, such as global popularity of items and associations among items. The first recommendations are predefined, while the next ones can be adjusted to the user’s actions by a session-based recommendation algorithm. (c) The need for the RS to be dynamically adjusted by exploiting interactivity/filtering. The problem in this case is not simply to produce a list of items but involves the ability to react with intelligence to the user’s selections.
• Personalizing promotional actions. The problem is to match to a particular user any type of action, such as to send info about a product, show an ad, make a special offer/discount, send coupons, etc., on behalf of stores, either online, offline, or both. The RS may be operating in (a) search engines or metasellers (eBay, BestPrice, etc.) or (b) individual stores, where the RS is typically incorporated into their loyalty program and used to send promotional messages, discounts, gifts, etc. A rapidly growing trend is for companies to offer to their customers mobile apps linked to a customer account. Such apps are inherently personalized bidirectional channels and one of the most promising fields for RS.

The focus of research in e-commerce [9] falls into the broad areas of (1) understanding consumers (collect information and build user profiles), (2) delivering recommendations (match users with items and produce recommendation lists, and (3) recommendation system impacts (maximize business impact, such as user acceptance, satisfaction, and loyalty by considering issues such as timing of delivery, frequency and number of recommendations, provision of justification, and presentation issues). Recommendation problems can thus be broken down to challenges related to any or all of those areas. In the following subsection, we briefly review the main recommendation approaches, depicting the special characteristics of each one and the problems where it fits best, while in the next subsection, we review the core methods and related research efforts in RS for retail.

5.2. Outline of Main Recommendation Approaches Applied in e-Commerce

The core of every RS is the algorithm responsible for matching users with items. A variety of recommendation approaches exist, some of which are more simple, while others may combine several technologies in a complex structure. Figure 2 summarizes the main advantages and disadvantages, as well as the most typical technologies applied in each recommendation approach.

5.2.1. Collaborative Filtering (CF)

Collaborative Filtering can be characterized as the most typical recommendation approach, where large volumes of historical data, such as purchases and evaluations, are used to make recommendations exclusively based on users-to-items relationships. We distinguish two cases, user-based and item-based CF. User-based algorithms find for the target user a set of neighboring ones, using some “user similarity” measure. They then aggregate the items purchased or positively evaluated by the most similar users, eliminate those items that the target user has already purchased, and recommend the remaining items. In item-based CF, instead of finding the preferences of similar users, the RS finds items that are similar to the ones for which the target user has expressed preference in the past. The item-based technique is often used to calculate recommendations for big online shops, where the number of users is much higher than the number of items.

In CF, we distinguish memory-based algorithms, which utilize the entire user–item database to generate a prediction, and model-based approaches, where a learning algorithm constructs a model of the user–item interactions. In e-commerce, where the problem is to identify suitable items, the most common approaches for model-based CF are clustering, classification, and Markov decision process (MDP). These algorithms are able to solve the shortcomings of memory-based CF algorithms by drastically reducing the memory and processing power requirements. Additionally, the produced model can be seen as a solid and interpretable model of the customer behavior, and can be used for product positioning and tuning of marketing RS [41]. It is noted that other CF methods, such as regression models and SVD/matrix factorization, are more suitable to the prediction of quantitative ratings, typically found in content recommendations (e.g., films and books).

The strength of CF is that it does not require any knowledge of the properties of the items themselves, which makes it very practical in cases where such knowledge is not available or costly to create and maintain. On the other hand, it suffers from the cold start problem (when there are not yet enough usage data to match all users with all items) and the ramp-up problem (when new unknown items are frequently added) [42].

5.2.2. Content-Based (CBF)

In content-based RS, recommendations are produced by matching a description of an item with a profile of the user’s interests. There are three important aspects to a CBF recommender, namely, “the matcher”, “the item descriptions”, and “the user profile”. “The matcher” can be a rule-based algorithm or stem from the domain of information retrieval theory. The user profile is (a) a model of the user’s preferences, i.e., a description of the item types that interest the user and (b) the history of the user’s interactions with the e-commerce system, e.g., whether the user has purchased the item or has searched in the relevant category. The user model can be seen as a classification model, which predicts whether an individual user will like an item. The latter is the classification variable, while the item descriptions that a user has bought or rated are used as independent variables, and the available history of user actions is used as training data [43].

Content-based methods are advantageous in making recommendations for new items, especially when sufficient past user actions are not available for that item. On the other hand, an important disadvantage is that they tend to provide obvious recommendations with reduced diversity, since the constructed model is specific to the user at hand and the community knowledge from similar users is not leveraged. Additionally, even though content-based methods are effective at providing recommendations for new items, they are not effective at providing recommendations for new users. This is because an accurate user model needs sufficient information on the specific user’s history. Content-based RSs are considered to be closely related to knowledge-based RSs, since the similarity metrics for the matching are often based on domain knowledge [44].

5.2.3. Knowledge-Based (KB)

In the knowledge-based approach, the matching between users and items is based on knowledge in the particular domain, rather than the history of interactions between users and items. The knowledge can be about the features that items may possess and how these features can serve the needs of particular users. It can be asserted by human experts or mined using knowledge extraction methods, and it is modeled and maintained in a knowledge base. The input to the RS is a description of the user’s needs or interests, while the response is inferred by a reasoning engine. The result can be either generic or personalized, depending on the available knowledge. Knowledge-based RSs are particularly fit to deal with the cold start problem and whenever the items are complex and highly customizable. Their strength is that the output is based on accumulated, maintainable, and human understandable prior knowledge, which integrates past experience with human expertise. Reasoning is also applied so that axioms and factual knowledge can lead to inferred knowledge.

5.2.4. Demographic (DF)

Demographic RSs aim at categorizing users based on information such as age, gender, profession, and nationality, assuming that they will have similar preferences. The key element of DF is that it creates categories of users with similar demographic characteristics, and tracks the aggregate buying behavior within each category. Recommendations for a new user are issued by first finding to which category he/she belongs and then by applying the aggregate buying preferences of previous users in that category. DF is performed using clustering techniques, where, for a given cluster C, its density represents the number of users in it and its radius is a measure of how demographically dissimilar they are. Then, the historical data on buying behavior or preferences of each user in C are used to associate with the cluster C an aggregate buying behavior [45]. DF is a simple, straightforward approach with a small cost, which, however, has limited capabilities for accurate recommendations and is not able to provide personalization.

5.2.5. Hybrid Recommendation Techniques (HR)

In hybrid recommendation, more than one method is combined in order to overcome the limitations of individual approaches and improve the RS performance [10]. HR is closely related to the field of ensemble analysis, in which the power of multiple types of machine learning algorithms is combined to create a more robust model. Most commonly, CF is combined with a CBF or KB technique to deal with the cold start and ramp-up problems [42]. However, the combination of methods varies with respect to the scenario, environment, or problem. A recent example of a hybrid RS for e-shops is the one presented by Walek and Fajmon [46], where a CF component is combined with a CB component and a KB component (fuzzy expert system).

Interesting applications of hybrid RS were found in e-retail. Yan et al. [47] presented an RS for fashion e-shops, where a data augmentation component applied visual feature extraction to attach additional tags to products. The augmented features were then fed into a knowledge graph component, which performed the user–item matching. Ma and Jiang [48] developed a secondary filtering mechanism, which combines the output of an initial algorithm with a measure of the user’s novelty feature. The metalevel algorithm classifies the items to optimize the recommendation list according to the character of the e-shop user.

Bandyopadhyay and Thakur [49] reported that they successfully combined CF with a neural network component to recommend products to students, while Rodrigues and Ferreira [50] proposed a simpler algorithm, combining clustering and association rule mining to recommend items on behalf of a perfumery chain. A hybrid system applied to physical retail shopping was also proposed by Choi et al. [51], who combined CF with a location-aware component that considers venues and events in order to produce personalized product recommendations in malls.

5.3. Modeling Methods Most Commonly Used in Retail

In this section, the modeling methods most commonly used in electronic and physical retail are reviewed. Short comments for each category and related papers are summarized in

Table 1. Modeling methods successfully applied in e-retail.

Methods	Comments	Related Papers
Rule-based and knowledge-based models	Methods with long history in RS, suitable for alleviating the cold start, sparse data, and ramp-up problems. Complex models can be used to combine pre-existing domain knowledge with discovered patterns and product features. Recent research utilized semantic web technologies and reasoning engines in order to enhance the intelligence/cognitive abilities.	[24,52,53,54,55,56,57,58,59,60,61,62]
Neural networks and deep learning	The most rapidly developing modeling approaches, appearing in a variety of different forms (e.g., deep networks, recurrent, embedding-based, etc.). Their main strengths are in learning complex relations and in capturing semantic, sequencing, and contextual information. Their limitations are their high demands for training data and computational resources. They are the preferred technologies for modeling multiple-step behavior and for extracting features from unstructured data.	[38,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98]
Markovian methods	Less popular methods that aim at capturing the sequences of user actions in click stream data and session-based RS. Recent research work was limited.	[99,100,101,102,103]
Graph databases	A relatively new and promising framework, which is efficient in capturing logical relations among users and items. It is suitable for both CF and KB approaches.	[104,105,106]
Matrix factorization	The most typical approach for early CF recommenders, which is still valid as a powerful core method for reducing the dimensionality problem. Recent research work in e-commerce RS was limited.	[71,107,108]
Natural language processing	Particularly useful methods for extracting implicit information about users’ stances, regarding both their personality/behavior and their opinion on specific products. Additionally, NLP has been used for extracting item features from their text description.	[109,110,111,112,113,114,115,116]

.

5.3.1. Rule-Based and Knowledge-Based Models

Rule-based RSs have been applied as a form of CF but also as CBF or a knowledge-based approach. Techniques were often based on adaptations of the association rule mining to the recommendation problem or on its extensions to more advanced versions. Combinations with other methods have also been reported, where the rule component was used to enhance the overall performance. Association rule mining, as proposed by Agrawal et al. [117], although initially conceived for the market basket problem, has been widely applied to RS for e-shops by reformulating the problem. In its simplest form, instead of focusing on the global user behavior that the discovered rules express, collaborative recommendation was achieved by selecting subsets of association rules for which the antecedent included an item known to be preferred by the target user. The top-k recommendations were then the consequent of the k rules with the highest confidence [52].

The above natural application of market basket analysis has been extended to consider ratings such as like, neutral, and dislike [43]. Additionally, association rules can be discovered with the same methods and datasets so that, instead of expressing relations among different items, they express relations among the preferences of different users. In this complementary approach, the preferences of a particular user can be predicted from the preferences of other users. Furthermore, the rules can be generalized so that the antecedents are not restricted to items (or pseudo-items) but can express the presence of any user characteristic. In this way, rules are used to match user profiles with items, thus serving content-based recommendation. The advantages of association rule mining include its efficiency/scalability and the interpretability of the results. The latter can be important in applications such as e-shops, where it has been found that statements such as “you may like this because…” have an impact on the site’s conversion rate [33]. On the other hand, association rules have limited expressiveness since they can only relate the simultaneous presence of items of the same type. Some of the notable early efforts to improve rule-based RS were to perform multiple-level rule mining [53] by formulating higher-level rules that generalize lower-level ones (e.g., the rule $beer\Rightarrow crisps$ can be generalized to the rule $drink\Rightarrow snack$). Lin and Ruiz [54] developed an algorithm that used multiple support thresholds in order to improve the discoverability of rules for items of a different frequency, while Leung et al. [55] proposed a CF framework using fuzzy association rules and multilevel similarity.

Ghafari and Tjortjis [56] presented in their recent review paper considerable work on the enhancement of the association rule mining process using heuristics. Najafabadi et al. [57] combined association rule mining with clustering to deal with the data sparsity problem and enhance the performance of CF. They considered the number of repurchases of an item as an implicit measure of liking and used it to cluster user preference profiles. Association rules were then mined and applied in the enhanced clustered dataset. In their recent work, Pariserum Perumal et al. [58] refined the rule component by employing fuzzy logic so that the changes in users’ preferences over time are also considered. They showed success in coping with the dynamic change of users’ interests. Nair et al. [59] mined temporal association rules by developing a symbolic aggregate approximation method to convert time series to symbols. Liao and Chang [60] dealt with e-commerce recommendations with possible consumers’ behavioral changes by adapting association rules to the analytic hierarchy process (AHP), proposing rough-set-based rules.

An interesting rule-based approach in sustainable marketing was presented by Tomkins et al. [24]. The aim was to discover and recommend more choices of sustainable products in market segments where there is a lack of credible information and where alternatives are limited and difficult to find. The authors developed methods to identify sustainability-minded shoppers and then used their purchasing patterns to label the preferred products as sustainable, even if these products were not officially certified. Knowledge from purchasing patterns was then combined with domain knowledge and product metadata, using a probabilistic soft logic rule framework. The method was evaluated on the Amazon platform in food products, outperforming baseline NN-based and SVD-based methods.

A more advanced rule-based approach has been used in KB approaches by employing linked data and ontological rules. In this case, complex relations can be expressed in a problem-specific knowledge domain. An inference engine is then able to apply logic and to produce secondary knowledge in the form of inferred facts [61,62]. Kim et al. [61] presented how the ontology web language (OWL) can be used to formulate semantic rules and recommend apps to mobile users. Their approach was to capture attributes of products, the consumption history of users, and the social relations of users. Recommendation was driven by reasoning, which considered ontology-based semantic relations among apps and also determined the similarity between the target user and his social members. Aguilar et al. [62] differentiated the intelligent recommendation from knowledge-based approaches as those that, in addition to rules, employ learning, knowledge assertion, and reasoning. They proposed a framework for intelligent recommendation and implemented an RS using fuzzy cognitive maps.

5.3.2. Neural Networks and Deep Learning

Neural networks have attracted researchers due to their great capability in modeling complex relations among items and users, and are nowadays probably the most rapidly developing area. With respect to the number of incorporated layers, neural-network-based models can be divided into shallow and deep networks. Shallow networks include a small number of hidden layers, most often just one. Because of their simple structure, they are suitable for problems with limited complexity, but they train faster and can easily be applied on very large datasets. On the other hand, deep networks incorporate a large number of hidden layers and have a more composite structure, which gives them the ability to learn complex relations, including contextualized and time-varying processes. Their excessive computational requirements pose restrictions on the size of the datasets that they can handle. There is thus a trade-off, which makes shallow networks preferable for huge datasets, whenever the underlying relations can be satisfactorily represented by a nonlinear classifier, while deep learning is more suitable for capturing more complex relations, provided that there are sufficient computational resources.

Shallow Neural Models

A special type of shallow neural networks, which are the most relevant to RS, are those used in embedding techniques, such as word vector representations (word2vec) [63]. Inspired by the great success of word2vec in natural language processing (NLP), the embedding models in RS usually contain a shallow network structure [64], which maps the items within sessions or baskets into a latent space, in which it is more effective to capture inter-item relations. When mapping items into a multidimensional latent space, their positions reflect their relations so that the latent numerical vector representation of each item contains much richer information than the original item ID. These networks are also called “wide”, because they have just one hidden layer with a large number of neurons. The approach can be generalized to represent user profiles (e.g., user2vec) and thus capture user–user and user–item relations. In their more recent papers, Barkan et al. presented an advanced item2vec version, where the user representation includes dynamic characteristics derived from the history of the user’s actions [65], as well as a combination of item2vec with a hierarchical tag representation, in order to deal with the cold start problem [66].

A representative work, where a shallow neural model was used for embedding-based RS, was reported by Hu et al. [67]. Their effort was to improve personalized recommendations within user sessions by adding diversity and reducing the dependence of the result from the strict order of the user’s item selection sequence. Stalidis et al. [38] applied an embedding-based algorithm on e-grocery RS, comparing it with a statistical factor and clustering method. It was found that the NN approach achieved comparable precision with the statistical method and, at the same time, offered the ability to overcome the limitations of the latter to be applied on large-scale problems.

Deep Neural Networks

Deep neural networks (DNNs) are multilayer perceptrons with multiple hidden layers, initially proposed by Hinton et al. [68]. Hinton presented an unsupervised greedy layer-by-layer training algorithm, which solved the optimization problem related to deep structures. Deep learning is a relatively new field in machine learning research that has been proved particularly successful in RS. Other areas where it has shown great success are computer vision and automatic translation. In [118], the authors provided a review of the main deep learning approaches, in which, although not specialized to RS, it clarified the main types of deep neural networks, their features, and strengths. Considering that the most important drawback of DNNs is their excessive demand for computational power and memory, the interesting and highly cited work of Iandola et al. [69] focused on reducing the size of the deep learning models without significant loss in performance.

One of the most important abilities of DNNs is to learn complex high-order or multiple-level relations among items and users. They are good in learning from low-level features, either raw or combinatorial, such as sets of related products in a user’s interaction history or sequences of actions, to form more abstract high-level feature representations, e.g., user profiles or preferences for product attributes. The learned high-level features are then fed to subsequent components, which generalize user–item relations [70]. Deep learning is thus often used as an auxiliary component to extract information from raw data sources, which are difficult to process directly, such as free text, images, and video. In other research work, such as the neural collaborative filtering proposed by He et al. [71], a neural network is also used in the core of a CF setting, i.e., to represent the interactions among users and items, replacing the traditional matrix factorization techniques. Architectures based on deep networks are also found in content-based RS to enable the codification of more complex abstractions as data representations in the higher layers [72]. The NN captures the intricate relationships hidden in low-level data sources, such as visual, textual, and contextual information, e.g., finds visual similarities between products or detects the user’s style preferences from free-text comments.

A deep learning model most commonly used in e-commerce RS is the recurrent neural network (RNN) [73], in which, rather than considering the recommendation problem as static, it incorporates transition information by learning user actions as sequences of interdependent steps. The long short-term memory network (LSTM) is a special RNN architecture with additional memory components, which enable it to capture time dependencies in multiple time scales. Representative research work relying on the abilities of RNNs was presented by Lee et al. [74]. The authors aimed at a RS that would be suitable for market areas such as e-grocery, where repetitive purchases of the same items are common, but also significant changes in customer needs and preferences may occur in time. To this end, they used an RNN model to learn each user’s purchasing patterns as sequences and to recommend item sets that are both relevant and diverse in multiple time periods. AN experimental comparison of their model with a CF-based model showed that the prediction accuracy and the recommendation quality are considerably improved when the purchasing order is considered in markets where regular buys of the same products are common. On the other hand, RNN-based models assume that there is a strict sequential order in items, which may generate false dependencies. Salampasis et al. [75] compared the performance of RNNs and embedding-based RS in e-commerce, concluding that RNN achieved considerably better results, especially in predicting the next item to be recommended in shopping sessions.

The convolutional neural network (CNN) is another type of NN, which, although mainly applied to image and video recognition, has also been successful in RS [76]. CNNs are based on filters that slide along input features, transforming them into secondary feature maps. They are of relatively low complexity and thus more efficient than other deep architectures. However, they may lose some information due to the convolution operation, and they may be limited in capturing long-term dependencies due to the restriction on the size of their filters. CNNs are typically used for extracting features from unstructured data, such as item features (style, category, etc.) from images or semantics from natural text. A representative application of CNN is the research proposed by Addagarla and Amalanathan [77] to perform top-N recommendations in e-shopping platforms based on the visual similarity of products. They trained a CNN model in order to extract image features, and then generated image embeddings and built an index tree using the approximate nearest neighbors oh yeah (ANNOY) algorithm. According to their experiments, they outperformed other popular models, achieving an accuracy of 96.2%. In another recent paper, Latha and Rao [78] proposed an enhanced CNN model to analyze the customers’ sentiments on the Amazon product review database.

To deal with the challenges found in basic neural networks, several advanced NN models have been proposed: (a) attention mechanism, a technique for focusing on selective input parts [65,79]; (b) memory network, which captures the user and item interactions through incorporating an external memory matrix [80]; and (c) mixture models, which combine different models that can perform better in modeling sequential dependencies [81].

Xue et al. [82] focused on capturing higher-order item relations using a deep structure, while Cheng et al. [83] presented a wide and deep structure, which performed better than wide embedding-based networks and deep structures alone. Soon afterwards, Guo et al. [84] proposed a variation of the above wide and deep network called DeepFM. This included a factorization machine (FM) component and a deep component, which were simultaneously trained on a common raw-feature input. The DeepFM algorithm was able to capture low- and high-order feature interactions from raw features without the need for manual feature engineering. In Chen et al. [85], the authors proposed a transformer-based framework for the Alibaba recommendation system. They used the sequential behavior of each user as features, and they achieved state-of-the-art performance. Another successful combination of multiple NN components is the architecture proposed by Khan et al. [86], in which they applied a CNN for extracting contextual information from textual item descriptions, along with a W2V component for representing items and users. These modules were integrated with a CF component to provide top-N recommendations. In comparison with several competitive methods, the proposed model provided improved top-N item recommendation accuracy on the Yelp dataset.

Graph Neural Networks

In modern e-commerce RS, it is important to accurately recognize the different types of relationships in which products may participate, such as complements or substitutes, in order to generate recommendations with improved impact and explainability [87]. For example, substitutable items are interchangeable and can be proposed to increase diversity, while complementary items may be purchased together and can be recommended for up-selling. Such information about products and their associated relationships naturally forms product graphs, which can be exploited by graph neural networks (GNNs) [88]. GNNs have shown great expressive power in modeling complex relations by introducing deep neural networks into graph data and are widely applied for knowledge graph-based recommendation and social recommendation [89,90,91,92] and are even adapted to traditional recommendation methods, such as CF [93,94].

In Gao et al. [119], the authors review the literature on GNN-based recommender systems, discuss their strengths and challenges, and provide an overview of existing research in RS applications. Recently, Fan et al. [90] provided a principled GNN approach with social connections and user purchase history to capture the interactions between users and items. Song et al. [89] used a dynamic graph attention network and incorporated recurrent neural networks for user behaviors in session-based social recommendation. Grad-Gyenge et al. [95] built a graph embedding method that took advantage of the knowledge graph to map users and items for recommendation. Considering the user—item interaction, Wang et al. [93] constructed a user—item interaction bipartite graph and proposed a graph-based CF method to capture higher-order connectivity in the user—item interactions. Another interesting paper in this area is the work of Xu et al. [120], in which they propose their RElation-aware CO-attentive GCN model to effectively aggregate heterogeneous features in a heterogeneous information network (HIN).

5.3.3. Markovian Methods

A Markov chain is a model representing sequences of random variables and the probabilities of their states, each of which can take values from some set, such as actions in a commercial site. Markov proposed that the outcome of a given experiment can affect the outcome of the next experiment [121]. This type of process is called Markov chain. Hidden Markov models (HMMs) are a way of relating a sequence of observations to a sequence of hidden states that explain the observations.

Markovian methods can be used for modeling time directionality and logical dependence; i.e., user action predicts the interest in an item, rather than the other way around. For example, a user who has purchased a particular camera is likely to be interested in a matching accessory; however, the interest in an accessory does not justify the recommendation of the camera. In the early work of Eirinaki et al. [99], the training data were used to calculate the transition probability over a sequence of items. A user’s shopping sequence was then matched to the calculated sequence, and the transition probabilities were used for recommendation by selecting the candidate items with the highest probability. An HMM model was used by Shani et al. [100], who proposed an RS based on the Markov decision process (MDP). The states of the MDP were k-tuples of items, the actions corresponded to the recommendation of an item, and the rewards to the benefit from selling an item, e.g., the net profit. The state following each recommendation was the user’s response, such as selecting the recommended item, selecting a nonrecommended item, or nothing. A variation of the basic Markov-chain-based RS was proposed by Zhang and Nasraoui [101],, who combined a first- and second-order Markov model to make more accurate web recommendations. In their more recent work, Le et al. [102] developed a hidden Markov-model-based probabilistic model for next-item recommendations by incorporating additional factors like context features to leverage the recommendation accuracy. Another important variant was to adopt a factorization method on the transition probability to estimate the unobserved transitions [103].

5.3.4. Graph Database Modeling

The use of graph databases (GDBs) is one of the newest approaches to RS engine modeling. In a GDB, data are represented by graphs and stored using nodes, edges, and properties (or attributes). A graph database is a data model that focuses on entities and the relations between them. Rule-based recommendation systems can be implemented using graph databases. Konno et al. [104] modeled a recommendation system based on data-driven rules. They applied a two-layer approach to retail transaction data, where the knowledge graph database included a concept layer for semantic ontology representation and an instance layer for associating sets of retail business data to concept nodes. Using the Neo4j graph database [122], a clustering step was adopted to group customers into communities using RFM (recency, frequency, monetary) analysis [123]. Then, a list of recommended products was created for each group with a reasoning engine. The performance of the system in terms of time efficiency and the novelty of recommendations was found to be reasonably good. Another rule-based recommendation approach is described in Sen et al. [105], where the Neo4j tool was used for data modeling and product recommendation purposes. Raw text data were modeled on an Neo4j graph database with Cypher queries to build a graph data model. The proposed model captures the influence of a product on another product so that if a user bought the influential product, then the influenced products can be recommended by the system to the users. The authors demonstrated an RS based on the graph data model and compared it with the Apriori algorithm.

A graph-based solution developed using the Neo4j graph database was presented by Delianidi et al. [124]. In this paper, the authors focused on the efficient operation of the next-item recommendations for an e-commerce retail store. With the appropriate data modeling, by defining nodes and relationships between the nodes and by executing Cypher queries, the system first learns the pairs of co-occurring products that appear in the same sessions and calculates their degree of similarity. The next product recommendation is then derived from the similar products co-occurring with the product that the user is currently viewing.

5.3.5. Matrix Factorization

Matrix factorization (MF) is the most representative technique in CF and probably the most widely adopted in early recommenders. It fits very well to the prediction of user ratings, dealing with the limitations posed by sparse data. Several methods have been developed for performing matrix factorization, the most typical of which are singular value decomposition (SVD), principal component analysis (PCA), probabilistic matrix factorization (PMF), and non-negative matrix factorization (NMF) [107]. The basic idea in all cases is to transform the sparse $users\times items$ matrix into a latent factor space of low dimensionality, where user–item interactions are represented by dense vectors, with minimum loss of information. Recent research in e-commerce RS that is based on MF, at least at its traditional form, is relatively limited, since it has moved towards neural networks that perform better in representing user–item interactions and are not limited to linear relations [71]. As an example of recent work based on MF, Le et al. [108] proposed an advanced factorization machine that introduces the notion of basket and is aiming at detecting the customer’s latent intentions. The model then predicts the items that the customer may still need to complete his/her shopping goal and adjusts the recommendations according to the items already in the basket.

5.3.6. Natural Language Processing Methods

Most of the e-shops, as well as search engines, are nowadays collecting customer reviews that contain valuable information about customers’ needs and preferences. RSs can benefit by extracting sentiment and semantic information from such natural text. Recommendation models based on natural language processing (NLP) may have various approaches: sentiment analysis, word representation with embeddings, topic detection, keyword extraction, and in general any task that makes use of a free text description. Srifi et al. [109] presented research efforts where CF techniques were applied on customer review texts. The technologies used for NLP mostly involve shallow or deep neural networks, and the approaches followed can be categorized as those based on words, on topics, or on review sentiment. Representative recent papers on e-shopping RSs that employ NLP are:

In Tarnowska and Ras [110], the authors proposed a customer loyalty improvement RS employing an unstructured database. After transforming the unstructured text to structured, they applied sentiment analysis on the comments of their users to produce recommendations. In another study, Sharma et al. [111] built a system for product recommendations using the titles of the products as the text features and the word2vec methodology for creating embeddings for the similarity between them. Shoja and Tabrizi [112], showed that the insights from the customers’ text reviews are crucial for RS. They applied the latent Dirichlet allocation method and association rules to extract knowledge from text. In a similar research work [113], the authors implemented a fuzzy-logic-based product recommendation system using sentiment analysis and ontologies. They showed that the negativity of text reviews affects significantly the users’ behavior.

The current state of the art in incorporating sentiment analysis, at least as regards the technologies used, is the methodology proposed by Karn et al. [114]. The authors developed a sophisticated hybrid sentiment analysis component, employing deep neural networks (Bert-MCARU-GP-SoftMax model). This model was combined with a hybrid RS engine, which consisted of a CF component employing support vector machines (SVMs), a CB component based on item text descriptions and the term frequency-inverse document frequency (TF-IDF) algorithm, and a deep NN decision component employing the growing hierarchical self-organizing map (GHSOM). By comparing their methods with four competitive methods on an Amazon fine food dataset, they established that through the incorporation into conventional RS of additional information extracted from user reviews, the performance of personalized recommendations to e-shop customers is considerably improved.

6. Solutions to Specialized Recommendation Problems

A review of notable research efforts is presented, where the researchers refined or combined the RS methods outlined in the previous section to deal with more specialized problems in the particular domain of e-commerce and retail.

6.1. Context-Aware Recommendations

In context-aware recommendation systems (CARS), the system considers special circumstances or is adapted to a specific time period, when the targeted customer is expected to have special needs, e.g., marriage, vacations, or celebrations. Contextual information could include time, location, emotions, or activities, very often drawn from social media data. For example, the types of accessories recommended by a retailer might depend on both the season and the intention of the customer to participate in a special activity, such as holiday travel. It has generally been observed that the use of such contextual information can greatly improve the effectiveness of the recommendation process in a wide range of product categories [125]. Huang and Zhou [126] investigated the importance of timing in mobile shopping personalization and studied the relations among different motivations of consumers and their responses during before-search and after-search usage.

The knowledge on which a CARS can be based may include the exact list of all the relevant factors, their structure, and their values or none of them. Depending on what the system knows, we can classify the knowledge into fully observable, partially observable, and unobservable [125]. In the fully observable case, the contextual factors are known explicitly, e.g., when the RS is aware that the relevant context for recommending a shirt is the time (day of week, season, holiday), purchasing purpose (own use or gift), and shopping companion, as well as the values of these factors at the recommendation instant. In partially observable CARS, only some of the information about the contextual factors is known explicitly, while in the unobservable case, the RS builds a latent predictive model to implicitly capture the context. Depending on whether contextual factors change over time, it is also possible to differentiate recommenders into static and dynamic. In dynamic contextual RS, factors, their structure, and values are allowed to adapt over time.

In their overview of context-aware algorithms, Raza and Ding [127] explained that CARS can be seen as an extension of traditional recommendation by adding the context dimensions. To deal with the resulting high dimensionality, three alternative approaches have been noted: (a) contextual prefiltering, where contextual information is used to restrict input data; (b) contextual postfiltering, where the identified items are reranked or filtered to adapt to the particular context; and (c) contextual modeling, i.e., to introduce all the dimensions in the core algorithm. Representative algorithms that are based on the latter approach are extensions of matrix factorization, found in several variations, including tensor factorization [128]. However, it is noted that these methods are more often found in rating prediction problems, rather than item selection. A more promising approach for recommending products in retail is to extend user-based methods, which focus on learning user profiles in various circumstances. Fang et al. [129] presented a method that combined the implicit estimation of user profiles with contextual information to recommend items on the user’s mobile during his/her physical shopping experience.

The integration of context information into an RS has been successfully achieved in many complex scenarios using deep learning methods. Recurrent neural networks (RNN) have been successfully applied in several sequential modeling tasks, where the deep learning component was mainly focused on how to effectively model the contextual information or to alleviate the data sparsity problem [96,97]. A novel model has also been proposed, named context-aware recurrent neural network (CA-RNN). Instead of using the constant input matrix and transition matrix found in conventional RNN models, CA-RNN employs adaptive context-specific input matrices and adaptive context-specific transition matrices. The adaptive input matrices capture the context, such as time, location, and weather, while the transition matrices represent the longitudinal user behavior [98].

6.2. Session-Based Recommendations

A session is a collection of actions or events that happened in a period of time, usually in the form of a sequence, such as the clicks when visiting an e-shop or the products put in a shopping basket in one visit [7,130]. A session-based recommendation system (SBRS) is aiming at learning the user’s preferences within a particular session, predicting the unknown part of the session or the future sessions, in order to react to users’ actions [131]. Such a system is based on modeling the relations within or between sessions and is particularly useful when the user is unknown, as in the case of anonymous e-shop visitors. Since the RS has no access to a user profile, it predicts the user interests by monitoring the user’s actions at the time of the visit. Alternatively, in the case of known users, by taking each session as the basic input unit, an SBRS is able to capture both a user’s short-term preference from the user’s recent actions and the preference dynamics from one session to another [132]. Many e-commerce RSs, particularly those of small retailers, handle sessions independently, even of the same user, since cookies and browser fingerprinting are not reliable enough to effectively track the user IDs of their visitors over a long period of time. Moreover, for a smaller e-commerce site, it is not expected to capture more than a few sessions for the same user [133,134]. In the most elementary cases, the method only considers the last selection of the user, ignoring the information of past clicks.

The most typical problems in SBRS for e-commerce are (a) next-item recommendation, where, at a particular step within a session (usually product viewing or adding in the shopping basket), the RS proposes one or several items by mainly modeling the intrasession dependency [135], and (b) next-basket recommendations, which predict the items that would occur in the next session by modeling both the intersession dependencies and the intrasession actions. Le et al. [108] proposed a method that first predicts the intentions of the user and then finds the items missing from the current basket to complete his/her goal. Compared with next-item recommendations, next-basket recommendations have received less attention. The problem of supporting sustainable marketing has also been addressed by Hwangbo and Kim [136], who developed a session-based technique for recommending fashion items that fulfill the needs of sustainability-sensitive customers (e.g., marked as fair-trade product, using eco-labeled or recycled materials, etc.). The approach proposed by the authors included a component that, in addition to sequences of user actions, considered sequences of item attributes, such as style and brand.

Deep learning approaches have been used in SBRS to integrate session clicks with content information for the clicked items [137]. Hidasi et al. [138] built an SBRS on anonymous session data by assuming a strict order over the interactions within each session to predict the next item to click. They introduced an advanced RNN unit, namely, a gated recurrent unit (GRU), which receives as learning input both the current state of the session (e.g., the currently selected item/category) and previous actions. An enhanced approach making use of data augmentation was proposed by Tan et al. [139], who elaborated on some variants of RNN-based SBRS. Jing and Smola [140] devised an SBRS on nonanonymous session data with order assumption inside sessions. On the other hand, Hu et al. [67] built an SBRS to recommend the next item to purchase without the order assumption inside sessions, employing embedding techniques for nonanonymous session data.

A GNN has been introduced by Wu et al. [88] to model the complex transitions within or between sessions. They first transferred a dataset containing multiple sessions to a graph by mapping each session into a chain, where each interaction was a node in the corresponding chain and an edge connected each pair of adjacent interactions. Then, the graph was imported into a GNN to learn an embedding for each node (interaction), encoding the complex transitions over the graph into the embeddings. Finally, the learned embeddings were imported into the prediction module. A graph database was also used in [124] for achieving efficient session-based next-item predictions in a retail store.

CNN is another good choice for SBRS for two reasons: (1) it relaxes the rigid order assumption over items within a session commonly used by RNN, which makes the model more robust; (2) it has high capacity in learning local features from a certain area and relationships between different areas, which can effectively capture the union-level collective dependency that is usually ignored by other models. The model proposed by Yuan et al. [141] contained an embedding layer, fully-connected layers, and both a horizontal and a vertical convolutional layer.

In Salampasis et al. [142], the authors reviewed the major tasks of SBRS—prediction of the next item, prediction of the next basket, and purchase intent. They also compared experimentally, in a real-world e-shop, a large range of methods, from simpler statistical co-occurrence methods to embeddings and deep learning. It was found that LSTMs consistently outperformed other methods of SBRS in all tasks, naturally modeling the dynamic browsing that happens in e-shops. It was also confirmed that the temporal focus of LSTM on recent behavior is better suited for item prediction tasks.

6.3. Group Recommendations

The addressed problem is cases where the products are consumed by groups of customers, rather than individual persons, which means that the RS should match varying or possibly contradicting preferences. Group recommendation is a popular problem mainly in the areas of entertainment and traveling. However, interesting usage scenarios have been reported in retail, such as food shopping for families. Such problems are more complex since, in addition to learning and matching the preferences of individual members, a strategy needs to be developed on how to aggregate the preferences of the group and to decide about the matches that optimize the overall satisfaction. The two main approaches are either to produce recommendations for individuals and then aggregate them into a group recommendation or to first aggregate individual preferences by constructing a group model and then to produce recommendations that match the group profile [143]. Several aggregation strategies have been proposed [144], such as average (averaging all individual ratings), least misery (taking the minimum of individual ratings), multiplicative utilitarian (multiplying the individual ratings), and average without misery (removing items anybody really hated and averaging the ratings of the rest). Trang Tran et al. reviewed group RS in the healthy food domain, where individual dietary needs, eating behavior, and taste were considered to recommend meal selections for families and other groups [145]. In the same domain, Berkovsky and Freyne [146] explored group recommendation strategies, weighting models, and heuristics. They proposed a hybrid approach applied together with CF algorithms, where the group aggregation component switched between aggregating recommendations and aggregating profiles. Park and Nam formulated the problem of optimally selecting the items that a physical store will display to attract the visitors of this store [147]. Their approach was to consider it as a group recommendation problem, where the potential customers of the specific store constitute a group and the overall goal is to propose to the store manager the items that optimize the overall appeal to the group.

Pessemier et al. [143] assessed the alternative aggregation strategies in combination with several recommendation approaches, evaluating the group recommendations in terms of not only accuracy but also quality metrics, such as diversity, coverage, and serendipity. Masthoff and Gatt [144] deepened the study of group satisfaction by considering how an individual member’s preferences may be influenced, not only by his own experiences, but also by the satisfaction or experiences of other members of a group. They modeled the concept of emotional contagion, i.e., a phenomenon where the affective state of a member is influenced by the affective state of other members, either positively or negatively. This effect differs according to the type of relationships within the group; e.g., in authority ranking relationships, the preferences of a member, such as the older or the most respected, may weigh more than those of others and propagate to the group. Another concept was the tendency for conformity, according to which an individual’s opinion tends to adjust to the opinion of the majority. Quijano-Sanchez et al. proposed a generic methodology supported by software tools to develop group recommenders using social elements, i.e., taking into account the relations among group members [148]. They showed the validity of their approach in instantiating a group recommender that runs as Facebook application and provides suggestions for shopping clothing, considering the opinions of other socially connected users.

A problem similar to group RS that is worth mentioning is to recommend combinations of products in the form of packages so that the overall liking of the user is maximized. In this case, instead of aggregating the preferences of a set of users for an item, the algorithm aggregates the preferences of a single user for a set of items. Beladev et al. proposed a method for product bundling [149], where, in addition to the packaging of products in bundles and the matching between product bundles and users, they consider the issue of pricing so that the overall profit for the e-shop is maximized.

6.4. Explainable Recommendation Systems

The explainable recommendation system (ERS) is another type of RS that, besides providing highly reasonable recommendations to the users, also focuses on notifying the reason behind the decision of recommending each item. In other words, an explainable recommendation tries to address the problem of “why” a user receives the specific proposals, and it aims at improving the transparency, persuasiveness, effectiveness, trustworthiness, and satisfaction of the RS [150]. In general, ERSs are recently being favored and heavily studied because they help developers in better understanding the reasoning behind the systems’ conclusions, while also effortlessly performing maintenance and debugging tasks.

The most common methods for achieving ERS are neural networks [79,150] and knowledge graphs [151,152,153], which can provide useful information regarding the reasoning of recommendation. Latent factor models (LFM) can utilize collaborative filtering and achieve satisfying results in terms of prediction accuracy but fail to clearly explain the results to users. Based on that, Zhang and Chen [150] proposed the explicit factor model that justifies that explainability and effectiveness are not two conflicting goals for an RS. It is expected that ERS will attain even larger focus by the scientific society because of the rise in users’ concerns regarding the monitoring and utilization of their personalized data and behavior. Walek and Fajmon [154] proposed another, much simpler approach, where a hybrid RS justifies the recommendations by stating whether they were based on their similarity with viewed items (produced by a CB module), on the user’s ratings (produced by a CF module), or on the customer’s purchase history.

6.5. RS Based on Implicit Information

RSs are collecting users’ preference data in two ways, either implicitly or explicitly. The latter means that users provide themselves their preferences and requirements, being directly asked by the system. Such type of data collection is the review, the rating, or the questionnaire regarding a product. The collection of implicit data is achieved by monitoring the users’ behavior, which is typically performed transparently. Such type of data can be the time spent on reading the item description or the click sequence while browsing an e-shop [155]. It is obvious that explicitly obtained data are more worthy, but the huge challenge of acquiring them has lead several research teams to focus on the exploitation of implicit data, which in many cases may be the only ones available or just easier to collect.

To deal with this challenge, Schoinas and Tjortjis [156] built a recommendation engine solely on multisource implicit feedback (IF). They highlighted that the common feedback of purchase history is not enough, since it limits the number of users and items that are available for recommendation. The absence of more accurate information (actions), such as view and search for a product, leads to reduced coverage. Addressing the same challenge in mobile channels, and in order to promote the sustainable development of e-businesses, Guo et al. [157] proposed an information-fusion approach. They developed an NN-based structure that fuses information from multiple sources, including the location of the user’s mobile device, activity in social media, sociodemographic characteristics, and attributes of searched products. The conclusion from this work was that information fusion can improve the accuracy of recommendations and the convenience of customers during their searches but, at the same time, faces data access restrictions and raises privacy concerns.

A simpler approach was developed by Choi et al. [51], who suggested a method for deriving implicit ratings of users on items from transaction data, as an alternative to explicit ratings. The equation of a user’s preference of an item is calculated from the number of transactions of the user that include the item and the total number of the user’s transactions. The IF features proposed in the literature are numerous, and there are several that are not obvious. Peska and Vojtas [158] explained the presentation context of features and utilized traveled mouse cursor distance, cursor in-motion time, scrolled distance, and time of scrolling. Unger et al. [97] proposed a context-aware RS, which infers the user’s activity by discovering patterns in the sensor signals of the user’s mobile device, such as motion, location, and ambient light. Utilizing a deep-learning-based factorization model, Xue et al. [159] merged explicit ratings with implicit feedback derived from the user’s actions showing nonpreference of items.

6.6. Social Networks and Trust-Based Recommendations

Despite their maturity, traditional approaches are all based on the assumption that users are independent of each other. However, users actually form friendships; they communicate through social media and influence each other so that certain users inspire more trust. Generally, trust-based RSs are based on user relations considering that a recommendation made by a trusted user has more value to the target user. Seckler et al. [160] noted how valuable is the trusted content for website users and how it can affect an RS. Several researchers have proposed RSs that employ trust and relation information captured in social media to optimize recommendations [115]. Li et al. [161] proposed a personalized product RS for retail e-shops, which combined three social analysis components, namely, preferential similarity, trust, and social relation analysis. Their experiments were performed in collaboration with volunteer customers of Yahoo Shopping. They reported high performance in personalized recommendations; however, input data were collected through direct questionnaires, rather than machine learning. The work by Shambour and Lu [116] concentrated on trust-semantic extraction between the users using collaborative filtering techniques for e-business applications.

A recent trend in social media marketing is that users are encouraged to become selling agents, promoting items through their social connections. Xu et al. [120] dealt with the learning of complex interactions among users, selling agents, and items within social networks by developing a relation-aware model based on a graph convolutional network (GCN), which aggregates heterogeneous features. In their experiments with real-world e-shop datasets, their relations-aware RS outperformed other baseline GCN- and MF-based systems.

7. Discussion

7.1. Summary of Research Landscape

The focus of the current review was on RS for retail shopping, including recommendations presented in e-shops or promotional messages that target customers of physical stores through loyalty programs. The particular field is becoming increasingly popular, following the explosive growth in e-commerce and mobile shopping. In addition to research regarding the core algorithms and the recommendation approach, there is significant and continuously rising research work on the business aspects of recommendation, i.e., marketing research in consumer behavior, consideration of user acceptance issues, and techniques for maximizing the impact towards the end business goals of the vendor. The problems addressed in recent years were more complex, beyond the performance of the core matching algorithm and issues such as cold start and scaling. Such challenges were to balance accuracy with discovery in order to best fit the user’s desire, to consider contextual factors, and to address the influence of trust and explainability.

Considerable progress has also been reported in exploiting multiple sources of information, such as the comments, links, and in general the behavior of users in social media. Several efforts focused on extracting implicit information from user actions and predicting their intentions, while other authors studied the impact of timing, presentation, and trust issues to maximize the acceptance on behalf of the users. In

Table 2. Marketing challenges addressed in the reviewed research.

Challenges	Methods and Proposed Solutions	Related Papers
Introduction of diversity, novelty, and other properties related to recommendation quality	Formulation of suitable metrics and development of algorithms that optimize multiobjective user–item matching functions. Detection of users’ expectations and personalization of results employing clustering, prefiltering, and item list reranking.	[22,25,27,28,29]
Optimization of user satisfaction and acceptance	User requirements and the effects of timing, overload, explanation, and other parameters have been studied through surveys.	[30,31,32,33,34]
Alleviate cold start and sparsity	The content-based and knowledge-based approaches are naturally stronger in this perspective. Various types of hybridization have been proposed, employing neural networks, advanced knowledge-based components, rule-based components, and data augmentation. In recent experimental comparisons, the combination of CF with DNN gave the best hit ratio.	[47,48,49,50,51,156,162]
Provide explanations for the recommendation	Knowledge-based systems were mainly used to provide the reasoning behind the recommendations. Knowledge graphs were successfully applied, as well as embeddings and specialized neural-network-based methods.	[150,151,152,153]
Operation with unknown users	Session-based RSs, being based solely on the current user actions, were applied to serve unregistered or rarely revisiting users. Most of the latest efforts employed deep learning networks.	[106,130,131,135,138,139,140]
Contextualization	The most typical approach was to extend matrix factorization (e.g., tensor factorization) to deal with the increased dimensionality. More elaborate solutions were based on specialized deep neural network architectures (e.g., context-aware recurrent neural network).	[86,96,97,98,125,126,127,128,129]
Implicit user information	Implicit feedback was derived from the user’s browsing sequence, transaction data, mouse movements, mobile device sensors, using hybrid approaches, and a variety of methods from association rules to deep learning.	[51,97,155,156,157,158,159]
Group recommendations	Several approaches were proposed, including weighting, aggregation of profiles, aggregation of item lists, and hybrids. Considerable efforts were focused on introducing user behavior and quality metrics.	[143,144,145,146,147,148,149]
Feature extraction from unstructured data	Information was successfully extracted from free text comments, image/video, and sound to enhance recommendation. NLP/sentiment analysis was used to exploit user comments and CNN components in deep networks to extract high-level features from visual or sound data.	[110,111,112,113,115,116,161]

, we summarize the most important challenges in the studied area, as found in the reviewed papers. As regards the technologies used, a shift was evident from traditional matrix factorization CF and rule-based KB systems to neural networks, especially to embedding-based techniques and deep structures. Despite their excessive computational requirements, they showed enormous potential in solving problems with increased complexity. Deep learning methods showed strong advantages, from which the most important in our view are (a) They improved automation and reduced the need for semiautomatic feature extraction. (b) They managed to successfully learn complex relations. Through architectures with multiple specialized components, it was possible to capture sequence patterns, to extract high-level features from low-level data (e.g., to generalize style preferences from visual input), and to understand semantic content. Moreover, they were able to mix such elements into highly powerful systems.

7.2. Remaining Challenges and Future Trends

The progress already visible in solving complex recommendation problems is expected to continue, enforced by two strong drivers: the increasing availability of huge volumes of e-commerce data and the strong processing capabilities of big data platforms. These factors are in favor of deep learning methods, which are expected to be more widely adopted. The upcoming goals, therefore, seem to be even more intelligent recommendations that more precisely adapt not only to users’ explicit needs and transactions but also to their personality and hidden desires. Marketing research is also progressing fast; however, the incorporation of its findings into RS is still limited. A direction for future research, where there are many challenges, is thus towards interdisciplinary approaches that merge the latest marketing techniques with state-of-the-art recommendation algorithms.

Considering the wide demand for effective RS in e-commerce by actually all modern commercial websites and vendors in retail, another anticipated challenge is the development of flexible RS that would be accessible to smaller enterprises with limited investments and would perform well in specialized niche markets. To this end, progress is envisaged in solutions that require minimum manual refinement; they can fit in existing e-commerce systems as low-cost components and work more intelligently even with sparse data in rapidly evolving inventories and market landscapes.