Enhanced Multi-Level Recommender System Using Turnover-Based Weighting for Predicting Regional Preferences

Abstract

In the realm of recommender systems, the prediction of diverse customer preferences has emerged as a compelling research challenge, particularly for multi-state business organizations operating across various geographical regions. Collaborative filtering, a widely utilized recommendation technique, has demonstrated its efficacy in sectors such as e-commerce, tourism, hotel management, and entertainment-based customer services. In the item-based collaborative filtering approach, users’ evaluations of purchased items are considered uniformly, without assigning weight to the participatory data sources and users’ ratings. This approach results in the ‘relevance problem’ when assessing the generated recommendations. In such scenarios, filtering collaborative patterns based on regional and local characteristics, while emphasizing the significance of branches and user ratings, could enhance the accuracy of recommendations. This paper introduces a turnover-based weighting model utilizing a big data processing framework to mine multi-level collaborative filtering patterns. The proposed weighting model assigns weights to participatory data sources based on the turnover cost of the branches, where turnover refers to the revenue generated through total business transactions conducted by the branch. Furthermore, the proposed big data framework eliminates the forced integration of branch data into a centralized repository and avoids the complexities associated with data movement. To validate the proposed work, experimental studies were conducted using a benchmarking dataset, namely the ‘Movie Lens Dataset’. The proposed approach uncovers multi-level collaborative pattern bases, including global, sub-global, and local levels, with improved predicted ratings compared with results generated by traditional recommender systems. The findings of the proposed approach would be highly beneficial to the strategic management of an interstate business organization, enabling them to leverage regional implications from user preferences.

1. Introduction

With the advancement of communication technologies and data acquisition strategies, substantial volumes of data from diverse sources have been collected by multi-state business organizations. These data encompass customer profiles, reviews and ratings, and purchasing frequency, among other elements. Such data sources offer significant opportunities for the strategic management of inter-state business organizations to extract valuable insights [1]. A data warehouse-based centralized analytical strategy has been widely recognized and employed for mining multiple data sources. However, challenges such as non-uniformity among data schemas, integration of disparate data sources, and the obliteration of unique data patterns present significant limitations. Consequently, the concept of a local pattern analytics strategy has garnered considerable attention for mining multiple data sources [2]. In this model, patterns from individual data sources are extracted and forwarded to the central level, rather than transferring the entire data source. This approach provides a two-level decision-making framework for both branch and central levels [3]. Locally mined patterns from branch data sources are instrumental in making local decisions, while global analytics of forwarded patterns are beneficial for central decisions. As a result, both local and global meaningful patterns are derived through the implementation of local pattern analytics. The schematic illustration of the local pattern analytics model is presented in Figure 1.

In recent years, there has been an increased focus on big data analytics for examining the multi-source data generated by inter-state business organizations. However, this field faces challenges that hinder both its theoretical and practical development [4,5,6]. Establishing a comprehensive big data framework for emerging applications still necessitates novel solutions [7]. The authors in [8] highlight the complexities inherent in big data analytics, describing it as involving diverse independent sources with distributed and decentralized control for composite and evolving representations. Consequently, applying a centralized computing approach (data warehouse-oriented) to the distributed and disparate nature of big data sources is not a feasible strategy for multi-state big data organizations. Furthermore, recent advancements in distributed technologies, open-source software initiatives, and the use of commodity hardware (commercially available hardware systems at lower cost) have further enhanced cost-effective solutions for implementing local pattern analytics. To reduce infrastructure costs, organizations are increasingly turning to commodity hardware and open-source software platforms and tools to meet their complex computational needs [9]. The Hadoop distributed processing framework, which includes HDFS (Hadoop Distributed File Systems) along with the Map-Reduce programming model [10], and the Spark distributed processing engine with in-memory based RDD (Resilient Distributed Dataset), are among the pioneering computational architectures for big data processing [11]. The Hadoop software Version 3.4 framework includes ecosystem software tools such as Hive (for SQL operations), Pig (for data pre-processing activities), Sqoop (for importing data from relational models), and Mahout (for implementing machine learning methods) to process big data sources without compromising the autonomy of data sources [12]. The Spark software Version 3.5 framework facilitates various ecosystem components, including Spark SQL for handling structured data, Spark-Streaming for processing stream data, and ‘MLib’ for implementing machine learning functionalities in a distributed manner.

As previously mentioned, the primary objective of local pattern analytics is to ascertain the interestingness of patterns at the data’s location without compromising the individuality of data sources. Consequently, big data processing frameworks utilizing ecosystem components reinforce the concept of local pattern analytics. The local pattern analytics model preserves the uniqueness of branches while concealing data complexities. Simultaneously, synthesizing novel patterns from a forwarded pattern base remains a challenging endeavor. In such contexts, weighting is considered a viable approach for aggregating and inferring information in fields such as probability and fuzzy set theory [13]. Accordingly, various weighting models have been proposed by researchers to synthesize local patterns from multiple data sources [8]. In [14], the issue of selecting data sources for multi-database applications is addressed. The work suggests applying threshold values for weighting the participatory data sources and pruning discovered patterns. The concept of heavy association rules is introduced in [15] to assign weights to the participatory data sources. A synthesizing model is developed to determine whether a heavy association rule is highly frequent or exceptional among participatory data sources. In [16], a weighting model for synthesizing high-frequency association rules from participating data sources is proposed. Here, the data source weight is calculated based on the number of transactions in the database. In [17], a weighting model for synthesizing high-frequency association rules is proposed in the context of multi-database mining. Here, the weight of a participating data source is calculated based on the number of high-frequency rules it supports. With the abundance of available information sources, it is imperative that organizations model complex customer-centric decision-making activities with increasingly personalized levels [18]. The ability to offer personalized recommendations makes the recommendation system an important pattern-driven technique among machine learning algorithms [19]. It has been integrated with state-of-the-art computing techniques [20] and has yielded promising results in various applications, such as recommending recipes [21], suggesting friends in social network communities [22], and others.

The remainder of this manuscript is organized as follows: the subsequent part of the current section details the implications of various research efforts conducted in the collaborative filtering-based recommendation system. The identified research gap and the research contributions are also presented in this section. Section 2 introduces various notions and terminologies used in the proposed approach. The schematic illustration of the proposed approach, along with three novel algorithms and illustrative examples, is presented in Section 3 of this manuscript. In Section 4, experimental investigations of the proposed approach are conducted on a well-known dataset, namely the ‘Movie Lens Dataset’, and the results are discussed. Section 5 of the manuscript outlines the concluding remarks and scope for future work.

1.1. Implication of Collaborative Filtering Technique

The fundamental concept underlying recommendation systems is the prediction of user–item interaction data. This approach leverages user-provided rating information to identify individuals with similar tastes or preferences and predict their interests [23]. Such systems not only assist consumers in acquiring desired products but also enhance organizational profitability by increasing sales and mitigating the issue of information overload when searching for user preferences. In [24], the authors proposed a decision support system for e-commerce enterprises to generate optimal collaborative suggestions based on purchase data. This framework provides various indicators for improved recommendation algorithms by analyzing customer shopping data as a sparse dataset. A novel recommendation system framework, termed ‘K-RecSys’, is introduced in [25]. This framework extends the item-based recommendation algorithm to incorporate domain-specific features of the fashion retail sector. The analysis of internet clickstream and offline purchase data is employed to ascertain consumer preferences. The authors of [26] offered a strategy for recommending human resources based on increased frequent item set mining. In [27], a customized product frequency recommendation system was developed using the ‘k-nearest neighbors’ algorithm.

Graph neural networks (GNNs) are also employed to model the intricate relationships among user behaviours within the context of recommendation systems. This approach addresses the challenge of data sparsity and enhances recommendation accuracy. A collaborative filtering recommendation technique based on GNN has been proposed for community identification and the creation of user and item characteristics [28]. In [29], a GNN-based mobile app rating prediction system was developed utilizing a semantic meta-graph structure. Factors such as user reviews, app genres, and rating information are employed to construct a similarity matrix. Subsequently, a GNN with an attention-based mechanism is utilized to aggregate feature information from neighbouring nodes. The concept of federated learning has also been successfully applied in the domain of recommender systems to create more privacy-aware user recommendations. In a federated learning setup, each user is considered an individual client with their own local data. Users can locally train an item-based collaborative filtering model, and model parameters are transmitted to a central server for constructing a user/item similarity matrix. The aggregated model provides a global perspective on user preferences. In [30], an efficient federated item similarity model for heterogeneous recommendation, termed FedIS, has been proposed. A two-stage perturbation method is employed in the system to protect both local training and transmission parameters, thereby enhancing privacy.

The implementation of a hierarchical approach in recommender systems facilitates the provision of more personalized recommendations across multiple levels by taking into account various tiers of user preferences and item characteristics. In [31], a framework for a multi-level recommender system tailored for tourism applications has been developed to aid potential travellers in identifying destinations that align most closely with their preferences and requirements. Scientific research in multimedia content analysis has indeed led to the creation of effective content-based recommender systems, as evidenced in applications such as movie recommendations [32], innovations in film and television art [33], and multimedia content recommendations [34]. The authors of [35] proposed a recommendation system that utilizes genre and user profile data, including age, gender, employment, and region. Patterns are discerned by correlating individual profile information with user ratings. The multi-objective next-basket recommendation (MONBR) algorithm [36] has been devised to enhance the quality of recommendations by assessing the temporal relevance of items to users based on factors such as utility, popularity, stability, frequency, occupancy, and novelty. The collaborative filtering technique significantly promotes the generation and utilization of customized recommendations. Consequently, a naive Bayes-based user-centric collaborative filtering recommendation system has also been developed [37]. A blockchain-assisted collaborative service recommendation method for data encryption is presented in [38], employing a ciphertext policy attribute-based encryption algorithm to ensure confidentiality and secure data exchange. The authors in [39] proposed a probabilistic matrix factorization-based recommendation approach that incorporates geographical location in computing users’ preference degrees. A customized learning resource recommendation technique based on a dynamic collaborative filtering algorithm has been proposed in [40], demonstrating a higher degree of relevance matching. The authors in [41] have investigated the collaborative filtering recommendation of online learning resources based on a knowledge association model. However, a trade-off between accuracy and efficiency persists in recommender systems.

1.2. Identified Research Gap

(i)

The collaborative filtering approaches previously examined predominantly rely on overall ratings, which are single-criterion in nature. However, utilizing customer ratings as a comprehensive evaluation strategy is inadequate for generalizing across all branches.

(ii)

As business operations are distributed across various geographical locations, conventional recommendation systems encounter relevance issues. This pertains to the alignment between user preferences and the items suggested by the model. Consequently, inter-state business organizations necessitate innovative strategies to perform sophisticated recommendations at multiple levels of abstraction with regional implications. This aspect remains unexplored in the existing literature.

(iii)

Furthermore, the evaluation of generated recommendations at appropriate taxonomic levels requires further attention. The sequence in which items are recommended based on local features of participatory data sources has not been addressed in the aforementioned research works.

1.3. Research Contribution

To address the identified research gaps, the contributions of this paper are summarized in terms of motivation, hypothesis, and technical execution.

(i)

Motivation: This study is motivated by the need to incorporate the significance of participatory branch-level data sources in identifying user preferences and computing predicted item ratings by users. The resulting weighted predicted ratings reflect the importance of both participatory data sources and user preferences, facilitating the exploration of recommendations at multiple levels of abstraction.

(ii)

Hypothesis: A turnover-based weighting model for participatory data sources is proposed to assign weights to branch data sources. Turnover refers to the revenue generated through total business transactions conducted by the branch. Thus, user ratings can be weighted and evaluated based on branch sales. Additionally, it is logical to establish a voting threshold for participatory data sources to exclude items preferred by a limited number of users. Users can rate purchased items on a scale of 1–5. The rating of the active user in each branch, who has purchased and rated the maximum number of items, can be utilized to calculate the weighted predicted rating. To enhance the predicted rating of recommended items with regional implications, it is appropriate to develop a multi-level recommendation system framework for global, sub-global, and local patterns. To substantiate the proposed weighted predicted rating method, three novel algorithms are introduced.

(iii)

Technical execution: A Hadoop-based big data processing framework is employed to technically execute the proposed approach. The Mahout machine learning ecosystem of Hadoop has been implemented to generate collaborative item-based recommendations in a distributed manner. To validate the proposed approach, experimental investigations are conducted on the Movie Lenz benchmarking dataset. The results clearly confirm the novelty of the proposed approach in calculating the predicted rating of items compared with the baseline technique.

2. Preliminary Terms in Proposed Approach

This section elucidates the terminologies and notations employed in the proposed weighting model. A summary of the notations utilized in this study, along with their respective descriptions, is provided in

Table 1. Summary of notations.

Notation	Description
BDi	Big data source of ithbranch or site
WBDi	Turnover-based weight of data source
TTD	Total turnover weight of data sources
W’BDi	Normalized weight of data source
TNW	Total normalized weight
Min.Ds	Standard minimum voting threshold
VBDi	Effective voting threshold of data source
U	Domain of all users
u	A user
Ru,i	Rating of user ‘u’ on item ‘i’
Ru,j	Rating of user ‘u’ on item ‘j’
${\overline{\mathrm{R}}}_{\mathrm{i}}$	Mean rating for item ‘i’
${\overline{\mathrm{R}}}_{\mathrm{j}}$	Mean rating for item ‘j’
Pu,i	Predicted rating of user ‘u’ on item ‘i’
aurj	Active user rating on item j
urj	Weighted predicted rating of user ‘u’ on item ‘j’
G-WPRj	Global weighted predicted rating of item j
SG-WPRj	Sub-global weighted predicted rating of item j
L-WPRj	Local weighted predicted rating of item j
F-LWPRj	Final local weighted predicted rating of item j
j∀sitei	For all data items J in site i
j∃sitei	Data items j existing in site i
j∃! Sitei	Data items j not existing in site i

.

2.1. Assigning Weights for the Participatory Data Sources

Let ‘BD₁′, ‘BD₂’… ‘BD_N’ be the multiple big data sources of ‘N’ branches or sites of an inter-state business organization. To infer the importance of participatory data sources, the weight of the data source in-turn assigned is based on the turnover cost of the respective branch. The turnover implies the total sales amount of the respective branch. Accordingly, ‘WBD₁’, ‘WBD₂’… ‘WBD_N’ are the turnover-based weights of data sources.

2.2. Normalized Data Source Weight

The normalized weight of a data source is calculated as the ratio between the turnover cost of the respective data source and the total turnover obtained from all participatory data sources. The normalized weight of W’BD_i is calculated using Equation (1).

$$Normalized weight of {BD}_i=W^{\prime }B{D}_i={\displaystyle \frac{W{BD}_i}{{\sum }_{i=1}^{i=N}W{BD}_i}}$$

(1)

The total normalized weight (TNW) is calculated by summing up the normalized weight of all data sources.

2.3. Estimation of Effective Voting Threshold for Participatory Branches

The effective voting threshold for participatory data sources plays a major role in the proposed turnover-based weighting model. For a given data source, the percentage of users who voted for each item can be calculated at first. Then the effective voting threshold is used to prune the items which are voted by the smaller number of users. A standard minimum voting threshold (Min.Ds) value has been uniformly chosen by the domain expert to compute the effective voting threshold of participatory data sources. The $V{BD}_i$ is denoted as the effective voting threshold.

$$V{BD}_i={\displaystyle \frac{Min.DS\ast \left(TNW-{BD}_i\right)}{1/N}}$$

(2)

2.4. Item Similarity Computation

The similarity between two items ‘i’ and ‘j’ is calculated using user’s rating on those items shown in Equation (3). To determine the similarity between the items, Pearson correlation similarity has been chosen.

$$\mathrm{P}\mathrm{e}\mathrm{a}\mathrm{r}\mathrm{s}\mathrm{o}\mathrm{n} \mathrm{S}\mathrm{i}\mathrm{m}\mathrm{i}\mathrm{l}\mathrm{a}\mathrm{r}\mathrm{i}\mathrm{t}\mathrm{y}\left(\mathrm{i},\mathrm{j}\right)={\displaystyle \frac{{\sum }_{u\in U}\left(R_{u,i}-{\overline{R}}_i\right)\left(R_{u,j}-{\overline{R}}_j\right)}{\sqrt{{\sum }_{u\in U}{(R_{u,i}-{\overline{R}}_i)}^2}\sqrt{{\sum }_{u\in U}{(R_{u,j}-{\overline{R}}_j)}^2}}}$$

(3)

2.5. Prediction of Rating

To predict the rating on the item for a target user, the weighted sum method has been chosen. It computes the prediction on item ‘i’ for user ‘u’ by computing the sum of the ratings given by the user for the item similar to item ‘i’ and each rating is weighted by the corresponding similarity between items ‘i’ and ‘j’. The prediction of rating by user ‘u’ on item ‘i’ is computed using Equation (4).

$$P_{u,i}={\displaystyle \frac{{\sum }_{allsimilaritems, N}\left(S_{i,N}\ast R_{u,N}\right)}{{\sum }_{allsimilaritems ,N}(\left|S_{i,N}\right|)}}$$

(4)

2.6. Active User Rating

The user can provide a rating for the purchased items on a scale of 1–5. The item-based collaborative filtering algorithms predict the rating of the non-purchased items for a user by computing item-based similarity. The active user for each branch can be identified by validating the following two constraints: (i) the user must purchase and rate a maximum number of items in the branch; (ii) the user predicting the rating for the target items should be a maximum of one.

2.7. Weighted Predicted Rating

Weighted predicted rating for the item ‘j’ can be calculated by multiplying the active user rating with the normalized weight of the data source where the item is found. The weighted predicted rating of item ‘j’(WPRj) is calculated using Equation (5).

$${WPR}_j\left(i\right)={\sum }_{i=1}^N{au}_{rj}\ast W^{\prime }B{D}_i$$

(5)

3. Illustration of Proposed Weighting Model

The schematic representation of the proposed weighting model is depicted in Figure 2. A layered approach is employed to illustrate the functionalities of the participating components. The data source layer encompasses the data sources of an inter-state business organization, comprising ‘N’ data sources, denoted as ‘BD1’, ‘BD2’, …, ‘BDN’, which are utilized to store customer transactions across ‘N’ branches. To process these data, commodity hardware, which refers to commercially available hardware, is employed in the subsequent layer, known as the storage and processing layer. The utilization of commodity hardware significantly reduces the cost of computing infrastructure while maintaining reasonable processing power. Each computing node is equipped with a Hadoop-based open-source software framework. This framework stores data in the native file system, specifically the Hadoop Distributed File Systems (HDFS), and processes it using a programming model called ‘Map-Reduce’. Additionally, one of the components of the Hadoop ecosystem, HBase, a well-known NoSQL database, is employed to store and process data in a scalable manner. Database-oriented read/write requests are communicated with the YARN (Yet Another Resource Negotiator) software framework of Hadoop 2.0, which is utilized for effective job scheduling and management of service requests. The subsequent layer, referred to as the analytical layer, is responsible for performing item-based collaborative filtering using the machine learning library of Hadoop, known as Mahout. This library employs a recommendation engine to predict the ratings of items found in the individual branch data sources. These predicted ratings are then forwarded to the next layer, termed the synthesizing layer. At this layer, weighting parameters, including the turnover cost of each branch, effective voting threshold, and effective voting rate of items, are applied, and the forwarded recommendations are synthesized into global, sub-global, and local patterns based on their presence in the participating branches. The top layer, known as the pattern discovery layer, is utilized to segregate and analyze the recommended patterns by the strategic management for further decision making. The rationale behind the proposed framework is to leverage big data processing layers within a commodity hardware setup to minimize operational costs and to discover regional patterns of interest with in-depth attention.

3.1. Algorithms Proposed

This section delineates the proposed algorithms incorporated within the framework. Algorithm 1 determines the effective voting threshold for each participating branch. Algorithm 2 calculates the voting rate of items by considering the number of users associated with individual data sources. Algorithm 3 computes the weighted predicted rating for globally, sub-globally, and locally recommended items.

Algorithm 1. Computing-effective-voting-threshold

Input: N—Participatory branch data sources WBDi—Weight corresponding to the turnover cost of branch data source BDi Min.Ds = Standard minimum voting threshold uniformly set by a domain expert Output:     The effective voting threshold for participatory sites Algorithm: Let TTD = 0; TNW = 0; Begin     For each data source i, i = 1 to N do     Begin     TTD = TTD + WBDi;     End;    For each data source i, i = 1 to N do    Begin      W’BDi = WBDi/TTD;      TNW = TNW + W’BDi;    End;     For each data source i, i = 1 to N do Begin              ${\mathrm{V}\mathrm{B}\mathrm{D}}_{\mathrm{i}}={\displaystyle \frac{\mathrm{M}\mathrm{i}\mathrm{n}.\mathrm{D}\mathrm{s}\ast (\mathrm{T}\mathrm{N}\mathrm{W}-{\mathrm{W}}^{\prime }\mathrm{B}{\mathrm{D}}_{\mathrm{i}})}{1/\mathrm{N}}}$     End; End all.

Algorithm 2: Voting-rate-item-count

Input:    N—Participatory branch data sources    u—No of users for individual branch data sources Output:    Voting rate of an item Algorithm:    Let rated-item-count = 0; voting-rate-item = 0.00;    Begin    For each data source i, i = 1 to N do      Begin         For each rated item j found in data source i;           rated-Item-count[j] = rated-item-count[j] + 1;           voting-rate-item[j] = rated-item-count[j]/u;            End;      Output: Voting rate of an item End all.

Algorithm 3: Turnover-based predicted rating

Input:         aurj—Active user rating for the item.           $\mathrm{V}\mathrm{B}\mathrm{D}$i—Effective voting threshold of participatory data sources.       Output:           Weighted Predicted rating.       Algorithm            Call computing-effective-voting-threshold ();             Call Voting-rate-item-count ();        Begin           For each data source i, i = 1 to m do         Begin            For each item j, j = 1 to m do           Begin               If $\mathrm{V}\mathrm{B}\mathrm{D}$i ≥ voting-rate-item[j] &&$\mathrm{j}\forall \mathrm{s}\mathrm{i}\mathrm{t}\mathrm{e}$i then               Begin                G-WPRj = ${\sum }_{\mathrm{i}=1}^{\mathrm{i}=\mathrm{N}}\left({\mathrm{a}\mathrm{u}\mathrm{r}}_{\mathrm{j}}\right)\ast$W’BDi;                 End;              else If $\mathrm{V}\mathrm{B}\mathrm{D}$i ≥ voting-rate-item[j] &&j∃site i then              Begin                SG-WPRj = ${\sum }_{\mathrm{i}=1}^{\mathrm{i}=\mathrm{N}}\left({\mathrm{a}\mathrm{u}\mathrm{r}}_{\mathrm{j}}\right)\ast$W’BDi;              End;             else If $\mathrm{V}\mathrm{B}\mathrm{D}$i ≥ voting-rate-item[j] && j ∃!site i then               Begin                 L-WPRj = $\left({\mathrm{a}\mathrm{u}\mathrm{r}}_{\mathrm{j}}\right)\ast$W’BDi;               F-LWPRJ = L-WPRj + ${\sum }_{\mathrm{i}=1}^{\mathrm{i}=\mathrm{N}}\left(\mathrm{u}{\mathrm{r}}_{\mathrm{j}}\right)\ast$W’BDi;             End;           else                Store item j as the uninteresting pattern for further investigation. End;       End;    End all.

3.2. Example

To elucidate the operational procedure of the proposed methodology, a user/item matrix has been constructed as an illustrative example, comprising fourteen users and twelve items. The ratings range from 1 (lowest rating) to 5 (highest rating). The MovieLens-100K dataset has been divided into three segments to demonstrate the participating branches. Additionally, the turnover cost of the branch data sources has been assigned to reflect the business volume of each branch. Accordingly, Site1 has been designated as a higher turnover branch (cost of USD 100,000), Site2 as a medium turnover branch (cost of USD 50,000), and Site3 as a lower turnover branch (cost of USD 20,000). The ratings provided by users for the considered items at the respective sites are presented in

Table 2. High turnover branch Site1.

Item/User	101	102	103	104	105	106	107	108	109	110	111	112
1	5	3	2.5	4		5		3	2	2.5	4	3
2		5		2	3		4	2	4	2	2.5	3
3		2.5	4		4.5	4.5	5			3	3.5	3
4		3		5	4	3				3	2	2
5		3	2	4	4	4						2.5
6		4	3		2	3	5		2		3
7	2	3	3		4.5	3.5	3.5	4		3		3.5
8	2			3	3	4.5		3.5			2.5	3
9			3.5	3.5						4		5
10		3.5	4			4	4	3		3.5	5	4
11	2	5			1	2			4.5	2.5	1.5
12	2.5		3			4	4.5			3	5
13		4	5				3.5	4.5		3.5		4

,

Table 3. Medium turnover branch Site2.

Item/User	101	102	103	104	105	106	107	108	109	110	111	112
1	5	3	2.5	4	5		3		2		4
2	5		2		3	4		2	4		2.5	3
3	2.5		4	4.5		4.5	5	5	3		3.5	3
4	3	4	5			3			3		2	2
5	4	3		4	4			4			2.5
6	4			3	2	5
7		2	3	3	4.5	3.5			4		3	3.5
8		2	3		3	4.5	3.5			2.5	3
9	3.5	3.5		5					4		5
10	3.5		4		4			4	2.5	3.5		4
11	2		5	1		2			4.5	2.5		1.5
12	2.5	3	3				4.5			3	5
13	4		5					3.5	4.5	3.5		4

and

Table 4. Low turnover branch Site3.

Item/User	101	102	103	104	105	106	107	108	109	110	111	112
1	5	3	2.5	4		5		3	2	2.5	4	3
2		5		2	3		4	2	4	2	2.5	3
3		2.5	4		4.5	4.5	5			3	3.5	3
4		3		5	4	3				3	2	2
5		3	2	4	4	4						2.5
6		4	3		2	3	5		2		3
7	2	3	3		4.5	3.5	3.5	4		3		3.5
8	2			3	3	4.5		3.5			2.5	3
9			3.5	3.5						4		5
10		3.5	4			4	4	3		3.5	5	4
11	2	5			1	2			4.5	2.5	1.5
12	2.5		3			4	4.5			3	5
13		4	5				3.5	4.5		3.5		4

, respectively. The predicted ratings, based on the item-based collaborative filtering algorithm [42], have been calculated. For instance, the rating for item ‘104’ by ‘User1’ is computed as 0.453 using Equation (3).

Similarity for the rest of the similar items is calculated as above. To predict the rating on an item for a target user, the weighted sum method has been applied. It computes the prediction on item ‘i’ for user, ‘u’ by computing the sum of the ratings given by the user on the item similar to item ‘i’ and each rating is weighted by the corresponding similarity S_i,j between items ‘i’ and ‘j’. The prediction P_u,i is computed using Equation (4). For P (User1,Item104) is 2.0. Hence the predicted rating on ‘104′ by ‘User1′ is 2.0. Similarly the predicted rating for all items is computed. By the proposed method, the normalized weights of participatory data sources are calculated using Equation (6).

$$\mathrm{Normalized} \mathrm{weight} \mathrm{of} {\mathrm{BD}}_{\mathrm{i}}={\mathrm{W}}^{\prime }\mathrm{B}{\mathrm{D}}_{\mathrm{i}}={\displaystyle \frac{{WBD}_i}{{\sum }_{i=1}^{i=N}{WBD}_i}}$$

(6)

• W’BD₁ = 100,000/170,000 = 0.5882.
• W’BD₂ = 50,000/170,000 = 0.2941.
• W’BD₃ = 20,000/170,000 = 0.11764.

The total normalized site weight is always 1 and the minimum voting threshold (Min.Ds) value has been uniformly set by the domain expert as 0.20 to compute the effective voting threshold using Equation (7).

$$\mathrm{E}\mathrm{s}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n} \mathrm{o}\mathrm{f} \mathrm{E}\mathrm{f}\mathrm{f}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e} Votingthresholdfor{BD}_i={\displaystyle \frac{Min.Ds\ast \left(TNW-{WBD}_i\right)}{\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$N$}\right.}}$$

(7)

• VBD₁ = 0.20 * (1 − 0.5882)/0.333 = 0.2471.
• VBD₂ = 0.20 * (1 − 0.2941)/0.333 = 0.42396.
• VBD₃ = 0.20 *(1 − 0.11764)/0.333 = 0.5299.

The voted rating of items found in the participatory sites is calculated based on the number of users who rated the items and the total number of users found in the sites. Accordingly, the voted rate of items is calculated and presented in

Table 5. Voted rating of items.

Item Id	Site1		Site2		Site3
	No. of User Voted	Voted Rating of Items	No. of User Voted	Voted Rating of Items	No. of User Voted	Voted Rating of Items
101	10	71.4	5	35.7	11	78.5
102	6	42.9	11	78.6	9	64.2
103	12	85.7	10	71.4	12	85.7
104	5	35.7	7	50	7	50
105	4	28.6	8	57.1	7	50
106	11	78.6	10	71.4	7	50
107	11	78.6	8	57.1	4	28.5
108	3	21.4	6	42.9	7	50
109	10	71.4	4	28.6	10	71.4
110	7	50	10	71.4	6	42.9
111	10	71.4	9	64.2	9	64.2
112	9	64.2	11	78.6	8	57.1

.

By comparing the voted rating of items and the effective voting threshold of sites, items have been classified as global, sub-global, and local. For example, Item ‘102′ scores the voted rating of 42.9 at Site1, which is greater than the effective voting threshold of Site1 (24.7). Its voted rating at Site2 is 78.6, which is greater than the effective voting threshold of Site2 (42.3). Similarly, at Site3, the voted rating is 64.2, which is greater than the effective voting threshold (52.9), and hence emerges as a global recommendation. Similarly, all the items have been classified into global, sub-global, and local based on their occurrences and resulted in

Table 6. Classification of recommendations.

Nature of Recommendation	Site Cluster	List of Items
Global recommended items	(Site1, Site2, Site3)	102, 103, 111, 112
Sub-global recommended items	(Site1, Site 2)	104, 105, 106, 107, 110
	(Site1, Site3)	101, 109
Local recommended items	(Site2)	108
Uninteresting items	(Site1)	108
	(Site2)	101, 109
	(Site3)	104, 105, 106, 107, 108

.

Now the weighted ratings are computed for the classified items by multiplying the rating given by the active user with the weight of the respective data source.

Global level weighted predicted rating (G—WPR_j).

$$\mathrm{G}\text{-}{\mathrm{WPR}}_{\mathrm{j}}={\sum }_{\mathrm{i}=1}^{\mathrm{i}=\mathrm{N}}\left({\mathrm{a}\mathrm{u}\mathrm{r}}_{\mathrm{j}}\right)\ast {\mathrm{W}}^{\prime }\mathrm{B}{\mathrm{D}}_{\mathrm{i}}$$

(8)

For Item 102 = 5 × 0.5882 + 0.2941 × 3.0 + 0.11764 × 2.3 = 4.09.

Similarly, weighted predicted ratings for other global recommended items are calculated.

Sub-global level weighted predicted rating (SG-WPR_j).

$$\mathrm{SG}\text{-}{\mathrm{WPR}}_{\mathrm{j}}={\sum }_{\mathrm{i}=1}^{\mathrm{i}=\mathrm{N}}\left({\mathrm{a}\mathrm{u}\mathrm{r}}_{\mathrm{j}}\right)\ast {\mathrm{W}}^{\prime }\mathrm{B}{\mathrm{D}}_{\mathrm{i}}$$

(9)

For Item 104 = 0.5882 × 2 + 0.2941 × 5 = 2.64.

Similarly, weighted predicted ratings for other sub-global recommended items are calculated.

Local level weighted predicted rating value (F-LWPR_j).

$$\mathrm{L}\text{-}{\mathrm{WPR}}_{\mathrm{j}}=\left({\mathrm{a}\mathrm{u}\mathrm{r}}_{\mathrm{j}}\right)\ast {\mathrm{W}}^{\prime }\mathrm{B}{\mathrm{D}}_{\mathrm{i}}$$

(10)

$${\mathrm{F}\text{-}\mathrm{LWPR}}_{\mathrm{J}}=\mathrm{L}\text{-}{\mathrm{WPR}}_{\mathrm{j}}+{\sum }_{\mathrm{i}=1}^{\mathrm{i}=\mathrm{N}}\left(\mathrm{u}{\mathrm{r}}_{\mathrm{j}}\right)\ast {\mathrm{W}}^{\prime }\mathrm{B}{\mathrm{D}}_{\mathrm{i}}$$

(11)

For Item 108 = 0.2941 × 3.4 = 0.999 + 0.5882 × 3.9 + 0.1176 × 3.8 = 3.74.

The calculated weighted predicted rating is tabulated and shown in

Table 7. Weighted predicted rating for recommended items for the example data.

Item	Weighted Predicted Rating
Global recommended items	[{102:4.09}, {103:3.4}, {111:3.5}, {112:3.5}]
Sub-global recommended items	[{104:2.64}, {105:3.38}, {106:3.38}, {107:3.76}, {110:3.56}, {101:2.82}, {109:3.27}]
Local recommended items	[{108:3.7}]

.

4. Experimental Investigations

This section details the experimental investigation conducted on the benchmark dataset to validate the efficacy of the proposed weighting model. The experimental setup was established using Hadoop version 3.2.4, in conjunction with the ecosystem component Mahout 14.01. As an open-source machine learning library, Mahout provides a comprehensive array of machine learning features and supports the distributed processing of large datasets across a cluster of nodes. To perform item-based collaborative filtering on the Hadoop cluster, the benchmark dataset, MovieLens100k [43], was utilized in our experiment. This dataset comprises 100,000 ratings (1–5) from 943 users on 1682 movies, with each user having rated at least 20 movies. For data processing, three Hadoop clusters were formed for the proposed study. In the ‘Site1 cluster’, 762 users rated 1590 items, with a total turnover cost of USD 100,000. In Site2, 942 users rated 1478 items, with a turnover cost of USD 700,000. Similarly, in the ‘Site3 cluster’, 927 users rated 1407 items, with a turnover cost of USD 600,000. To make effective branch-level decisions and understand customer preferences at individual sites, it is pertinent to also infer the predicted recommendations for non-purchased items. The proposed weighted item-based collaborative filtering algorithm was implemented across all three clusters, and the top 10 items of global, sub-global, local, and uninteresting nature were deduced from the participatory data sources. The weighted predicted ratings of the aforementioned classified items were calculated based on the proposed approach, and the items with their predicted ratings are presented in

Table 8. Proposed weighted predicted rating for recommended items.

Items	Site Cluster	Weighted Predicted Rating
Global recommended item	(Site1, Site2, Site3)	[{7:3}, {10:4.8}, {17:3.2}, {144:2.9}, {168:3.5}, {327:4.7}, {363:3.8}, {496:3.1}, {566:5}, {900:4}]
Sub-global recommended item	(Site1, Site2)	[{13:3.7}, {173:3.5}, {1039:2.0}]
	(Site1, Site3)	[{11:3.0}, {18:2.9}, {300:3.1}]
	(Site2, Site3)	[{286:2.7}, {659:1.9}, {1127:3.9} {1137:2.3}]
Local recommended item	(Site1)	[(14:3.4, {16:4.8}, {31:4.6}, {1763:3}, {1089:3.2}, {410:4.0}, {619:3.0}, {128:4.0}, {1098:2.5}, {1062:4.2}]
	(Site2)	[{114:4.1}, {188:4.7}, {220:3.9}, {270:3.5}, {303:4.0}, {411:2.0}, {686:2.5}, {882:3.0}, {1115:3.5}, {1963:2.0}]
	(Site3)	[{22:2.0}, {32:2.2}, {47:2.5}, {134:2.4}, {151:2.0}, {177:2.7}, {451:3.0}, {654:2.2}, {946:2.9}, {1009:3.0}]

.

In this study, a comparative analysis is conducted between the proposed approach and the traditional item-based collaborative filtering method to highlight the advantages of the proposed weighting model in calculating the predicted ratings of global, sub-global, and local items. Figure 3 illustrates the comparison of predicted ratings for the top 10 globally recommended items, which are endorsed by all participating sites, namely S1, S2, and S3. As these items are present across all participating sites, their predicted ratings are significantly influenced by the turnover-based weighting model, resulting in higher predicted ratings compared with the traditional approach. Figure 4 depicts the comparison of predicted ratings for items classified at the sub-global level. These sub-global recommendations do not appear in all participating sites but are endorsed by the majority. For instance, item 659 demonstrates that if all items are evaluated without assigned weights, the predicted rating of 1.9 would be absent, thereby failing to highlight the regional significance of such items in terms of their market potential to strategic management.

By employing the concept of data source weighting, the significance of regional patterns has been effectively captured, resulting in the proposed method achieving a superior predicted rating for all sub-global recommendations. Similarly, the effectiveness of the proposed weighted predicted rating method has been demonstrated for locally recommended items. These items are exclusively voted on within specific participatory data sources. Despite their infrequent occurrence, their implications are substantial for inter-state business organizations, particularly in terms of cost-benefit analysis. These local patterns are distinctive, with their marketability being high only on select platforms. Therefore, their predicted rating must be calculated by assigning an appropriate weight based on their participation relative to the turnover of the respective platform. The local recommended patterns, along with a comparison of predicted ratings in the participatory sites S1, are depicted in Figure 5.

In examining Item 1089 from Site1, it is evident that the discrepancy between the predicted ratings is substantial, with the proposed approach yielding a predicted rating of 3.2, compared with 2.0 using the traditional technique. If strategic management employs a threshold value of 2.5, the item would not be identified as an interesting pattern from Site1. However, under the proposed technique, the item attains its own significance and meets the threshold value. This suggests that the proposed weighting model offers an enhanced predicted rating and is capable of efficiently performing recommendations at multiple levels of abstraction, thereby segregating the discovered patterns based on global, sub-global, and local features.

4.1. Accuracy Evaluation and Baseline Comparison

To assess the efficacy of the proposed turnover-based weighted predicted rating approach, we conducted a comparative analysis with the traditional item-based collaborative filtering technique, henceforth referred to as the existing predicted rating method. This evaluation employed four standard accuracy metrics commonly utilized in recommendation system research: root mean square error (RMSE), mean absolute error (MAE), Precision@5, and Recall@5.

RMSE and MAE were calculated to determine the average deviation between actual user ratings and predicted ratings, with lower values in these metrics indicating greater predictive accuracy. The proposed weighted technique demonstrated a reduction in both RMSE and MAE, indicating its enhanced capability to more accurately predict user preferences. Precision@5 and Recall@5 were employed to evaluate the quality of the top-N recommendations generated by each method. Precision@5 measures the proportion of relevant items among the top five recommended items, while Recall@5 reflects the proportion of relevant items successfully retrieved among all relevant items for a user. The weighted approach achieved higher Precision@5 and Recall@5 scores compared with the traditional method, affirming its superior performance in identifying and recommending relevant items. The results are shown in

Table 9. Comparison of results in terms of standard measures.

Metric	Existing Predicted Rating	Weighted Predicted Rating
RMSE	0.926	0.782
MAE	0.734	0.602
Precision@5	0.602	0.710
Recall@5	0.518	0.621

.

The findings collectively suggest that the integration of turnover-based weights into the recommendation process not only enhances the accuracy of predicted ratings but also improves the relevance and efficacy of the generated recommendations. This highlights the practical utility of the proposed model in real-world applications where regional and economic factors significantly influence user preferences.

4.2. Sensitivity Analysis and Statistical Validation

To systematically evaluate the heuristic introduction of two critical hyper parameters—namely the linear revenue-based weighting and the minimum voting threshold (Min.Ds)—a comprehensive sensitivity analysis was conducted. Initially, the minimum voting threshold (Min.Ds) was varied incrementally from 0.10 to 0.30 in steps of 0.05 to assess its impact on prediction performance. The findings, depicted in Figure 6, reveal that both MAE and RMSE attain their lowest values near the originally selected threshold of 0.20, thereby empirically validating the expert-derived choice.

Subsequently, to assess the efficacy of the simple linear weighting based on branch turnover, a comparison was conducted with a logarithmic transformation of the revenue weight. As illustrated in Figure 7, the linear weighting scheme consistently results in marginally lower mean absolute error (MAE) and root mean square error (RMSE) compared with the logarithmic curve, thereby justifying the retention of the linear function for turnover-based weighting.

To ensure that the observed improvements are statistically robust and not merely a consequence of a single train/test split, five distinct random splits were generated and evaluated. Figure 8 illustrates the MAE and RMSE values across these splits, indicating that the performance of the proposed model remains consistent with minimal variation. This provides compelling evidence for the reliability and generalizability of the proposed weighting model.

In summary, these experiments substantiate that the selected hyper parameters and weighting strategy are well-founded, empirically validated, and contribute to consistent enhancements in prediction accuracy.

5. Conclusions and Future Directions

One of the issues with the present item-based collaborative filtering technique is that users and their evaluations are treated equally, with no regard or weightage given to the participatory data or ratings provided by the users. This paper introduces a weighting model based on the turnover cost of the participating sources and improved predicting rating is resulted. The proposed framework offers a sophisticated computational solution with the aid of commodity hardware and open-source software frameworks such as Hadoop and Spark. Hence the problems of data movement to the central repository and spending high investments on data warehousing-based centralized strategies have been eradicated. With the inclusion of the proposed weighting model, our intuitive framework performs personalized recommendations at multiple levels of abstraction and captures the regional behavior of users with improved accuracy. It has been proven with the aid of experimental investigations carried out on the benchmark dataset.

Assigning weights to participatory data sources based solely on a branch’s turnover cost may present certain limitations, particularly when the branch consistently experiences low sales volume. In such cases, alternative strategies, such as evaluating the number of patterns generated by the branch or correlating customer lifetime values with their ratings, may serve as viable measures for assigning weights. Beyond these static weighting strategies, weights can also be dynamically assigned based on factors such as frequency of visits, users’ purchasing history, location, and the characteristics of the items purchased. The adoption of such dynamic weighting schemes could enhance the accuracy of recommendations and enable strategic management to capitalize on unique patterns. Furthermore, the incorporation of advanced deep learning models in recommendation system research represents a promising future direction. Recurrent neural network (RNN)-based deep learning models, with their ability to process sequential and temporal data, are well-suited for modeling users’ temporal interests. Generative adversarial networks (GANs), another advanced deep learning model, can be employed to generate diverse recommendations due to their generator and discriminator functions. Additionally, GANs can effectively address the data sparsity problem when user ratings are unknown.