Discovering Key Successful Factors of Mobile Advertisements Using Feature Selection Approaches

Abstract

Programmatic buying has attracted growing interest from manufacturers and has become a driving force behind the growth of digital advertising. Among various formats, mobile advertisements (ads) have emerged as a preferred choice over traditional ones due to their advanced automation, adaptability, and cost-effectiveness. Despite their increasing adoption, academic research on mobile ads remains relatively limited. Unlike conventional statistical analysis techniques, the proposed feature selection methods eliminate the need for assumptions related to data properties such as independence, normal distribution, and constant variance in regression. Additionally, feature selection techniques have recently gained traction in big data analysis, addressing the limitations inherent in traditional statistical approaches. Consequently, this study aims to determine the key success factors of mobile ads in fostering customer loyalty, offering advertisers valuable insights for optimizing mobile ad design. This study begins by identifying potential factors influencing mobile advertising effectiveness. Then, it applies Support Vector Machine Recursive Feature Elimination (SVM-RFE), correlation-based selection, and consistency-based selection methods to determine the key drivers of customer retention. The findings reveal that “Price” and “Preference” are the most significant contributors to enhancing repurchase intention. Moreover, factors such as “Language”, “Perceived Usefulness”, “Interest”, “Mobile Device”, and “Informativeness” are also essential in maximizing the effectiveness of mobile advertising.

1. Introduction

Due to the COVID-19 pandemic, a significant portion of consumer behavior has shifted from in-store to online purchases [1]. As a result, many businesses have refocused their sales strategies on digital platforms. Programmatic buying is one of the leading developments supporting this transformation, which refers to using computer algorithms to deliver the right advertisement to the right consumer at the right time. In Taiwan, the programmatic buying market grew by over 30% in 2022, with 70% of advertisers adopting it and nearly half reporting satisfaction with the results [2]. Programmatic buying has thus become a driving force in the growth of digital advertising [3].

Programmatic advertising, a core component of programmatic buying, leverages big data and automated decision-making technologies to deliver highly personalized marketing messages. This includes real-time bidding, audience segmentation, and ad placement optimization [4]. In 2019, over 60% of the global digital advertising spend (USD 333.25 billion) was allocated to programmatic channels. Despite widespread adoption, academic research on programmatic advertising remains limited, especially within mobile contexts [5].

Simultaneously, the proliferation of mobile devices has dramatically changed consumer behavior and increased reliance on mobile commerce [6,7]. Mobile advertising—the delivery of promotional content through mobile platforms—has emerged as a dominant form of digital advertising [8]. It extends beyond traditional display ads to include formats on social media, video platforms, and native in-app environments [9]. The convergence of programmatic advertising with mobile platforms has introduced new opportunities and challenges in targeting, personalization, and user engagement [10].

Mobile advertising offers marketers a flexible, cost-effective way to reach users in real-time. However, consumer acceptance varies based on content relevance, perceived usefulness, and delivery context. Past research has emphasized informativeness, entertainment value, credibility, and user engagement [11,12,13,14,15]. While studies have explored these variables using traditional statistical methods, limited work has investigated their impact using feature selection approaches. Moreover, most existing models rely on assumptions such as data normality or independence, which may not hold in real-world marketing data [16,17,18].

Despite the increasing adoption of programmatic mobile advertising, academic research remains limited, particularly regarding theoretical integration. To strengthen the conceptual foundation, this study anchors key constructs in established theories: “Perceived Usefulness” and “Ease of Use” are derived from the Technology Acceptance Model (TAM); “Credibility” aligns with the Elaboration Likelihood Model (ELM); “Preference” and “Interest” reflect the Theory of Reasoned Action (TRA); and “Informativeness” is rooted in Uses and Gratifications Theory. These linkages enhance the theoretical contribution and contextual relevance of this study. Furthermore, few studies have systematically analyzed which mobile ad factors contribute most significantly to customer loyalty. Existing research often focuses on general attitudes or behavioral intentions rather than post-purchase loyalty metrics. Additionally, firm-level variables such as brand trust or competitive positioning are frequently excluded [19,20,21].

To address these gaps, this study aims to identify critical success factors in mobile advertising that enhance customer loyalty. We adopt three feature selection techniques—Support Vector Machine Recursive Feature Elimination (SVM-RFE), correlation-based selection, and consistency-based selection—to analyze survey data without relying on strict statistical assumptions. This approach helps uncover the most influential mobile ad attributes, offering theoretical insights and practical guidance for advertisers seeking to optimize engagement and loyalty outcomes.

2. Literature Review

2.1. Programmatic Advertising

Programmatic advertising has dominated online advertising, defined as “the use of data and technology that enables marketers to make decisions about the advertisement they want to send to the customers in real-time” [2]. It was projected to account for 86.12% of all display ad spending in 2022 [1] and continues to grow with automated data processing. It relies on real-time, data-driven automation to target audiences, often individual consumers. The Mobile Marketing Association defines mobile ads as “a form of advertising that can transmit advertising messages through mobile phones, personal digital assistants (PDAs), or other wireless communication devices” [22]. Compared to traditional campaigns, programmatic advertising uses advanced software to deliver highly targeted content based on user data. Its defining feature is automation, though definitions and interpretations vary.

Programmatic buying is a key driver of digital advertising growth [3]. Despite its importance, there is no unified definition. Shehu et al. [23] and Li et al. [24] describe it as a big-data-enabled IT practice that delivers personalized content via real-time bidding. This system lets advertisers bid for ad placements to reach the right audience at the right time [5]. However, relevant academic studies remain limited [25]. Concerns arise when premium ads appear on low-quality websites, negatively affecting brand perception [23]. Thus, this study focuses on app-based advertising. Despite the rapid rise of this multi-billion-dollar industry, programmatic advertising has received limited scholarly attention [26]. Meanwhile, apps have transformed digital behavior, enabling users to consume content and shop online, especially during the pandemic. Advertising has shifted from websites to apps, making mobile ads a significant revenue source for app developers [18]. Behavioral data allow for individual-level targeting, making data quality crucial. Advertisers must build detailed user databases to enhance targeting. This study, therefore, focuses on mobile ads to identify key factors that enhance customer loyalty.

2.2. Mobile Advertising

The mobile ads market in Asia reached USD 19.6 billion in 2021 [16], highlighting its growing importance. As defined by the American Marketing Association (AMA), advertising involves delivering persuasive messages to a target audience to promote products, services, or organizations [27]. It is a core marketing tool aimed at creating, communicating, and delivering value [27]. Mobile advertising extends this concept to mobile devices, using wireless technology to encourage purchases. Consumer acceptance is crucial. Guven [28] notes that while initial reactions to mobile ads may be negative, engaging content can increase tolerance. Similarly, Lu et al. [15] argue that relevance and perceived value are key to effectiveness.

According to the authors of [29], these factors reduce intrusiveness and positively influence attitudes toward ads. Maseeh et al. [30] distinguish between perceived benefits and risks, emphasizing that users accept mobile marketing when the value outweighs the drawbacks. Tan et al. [31] reinforce this, stating that acceptance depends on relevant content. Kerst et al. [32] further suggest that consumers compare an ad’s utility to that of alternative activities; higher utility leads to more favorable attitudes.

2.3. Customer Loyalty

Kwon et al. [33] describe brand loyalty as a firm commitment to repurchase or revisit a preferred product or service continuously, even when external factors or marketing strategies attempt to influence switching behavior. This definition underscores the two widely discussed dimensions of loyalty—attitudinal and behavioral [34]. According to Liu et al. [35], behavioral loyalty is reflected in repeated purchases of a brand, whereas attitudinal loyalty signifies an intrinsic commitment stemming from the perceived unique value of the brand. Therefore, in mobile advertising, customer loyalty is understood as a dual-dimensional construct that includes behavioral repetition and attitudinal dedication.

Drawing from prior research [36], customer loyalty in online shopping is a favorable disposition toward repurchasing, influenced by advertisements. It is considered a post-purchase phenomenon, in which consumers compare expectations shaped by mobile ads with their actual experiences of a product or service [37]. Customer loyalty is a key indicator of consumer behaviors such as repeat purchases, positive word-of-mouth, and brand commitment [38]. Within the Management Information Systems (MIS) field, customer loyalty is often used as a proxy for system success and has been extensively examined in empirical studies [39]. Accordingly, this study conceptualizes customer loyalty as a consumer’s overall emotional response based on their engagement with mobile advertisements from the m-company.

As noted by the authors of [40], customer loyalty is a key determinant of net benefits and individual impact. It drives repeat purchases and encourages consumers to recommend m-products or services to others. Shahzad et al. [41] further suggested that dissatisfied customers are more inclined to seek alternative options and may be more receptive to competitors’ marketing efforts than satisfied consumers. While previous studies indicate that mobile advertising is a strong predictor of customer loyalty [15], there is still a lack of comprehensive analysis that contextualizes this relationship. Therefore, identifying and defining the key factors in mobile advertising that enhance loyalty is essential. The following section explores these determinants. In addition, we will use repurchase intention to measure customer loyalty.

2.4. Potential Factors of Mobile Advertisements

This section identifies key factors influencing mobile advertising based on existing literature. Chen and Hsieh [22] outlined six essential elements for designing personalized mobile ads: pricing strategies, consumer preferences, promotional activities, user interests, brand perception, and mobile device type. These findings offer practical guidance for industry growth. Parreño et al. [42] highlighted entertainment value, irritation, and perceived usefulness as major drivers of positive consumer attitudes. Likewise, Flores et al. [43] found that engagement level, language, and ad placement context significantly influence consumer perception. Language-based ads enhance brand affinity, while placing ads on relevant platforms captures more attention [44].

Through a study on Korean consumers, Yang et al. [45] established that five elements—perceived usefulness, ease of use, entertainment value, annoyance, and credibility—are strong indicators of mobile ad effectiveness. Chen and Liu [44] highlighted concerns about data privacy regarding how individuals aged 18 to 24 perceive mobile advertisements. Meanwhile, Kim and Han [46] emphasized that informativeness, trustworthiness, entertainment, and incentives are central to encouraging users to engage with mobile ads. These four factors were further utilized to develop a predictive model for mobile ad effectiveness [33].

Additionally, Tseng and Wei [47] pointed out that mobile ads could adopt rich media formats, where users can engage with ads by expanding their display via a “Tap-to-Expand” function to enhance customer attraction [14]. Rich media mobile advertising includes diverse formats such as animated content, floating ads, interactive video ads, and expandable advertisements [15]. In Korea, cloud computing has enabled real-time, location-based, and context-aware formats [30]. With advancements in hardware and software—such as mobile technology and cloud computing—rich media mobile advertising has evolved to offer multiple presentation styles and targeted customer engagement strategies.

Chakraborty et al. [48] also observed that social media and game-based ad elements were widely incorporated among the ten most influential mobile advertising strategies. These allow direct consumer-brand interaction via platforms like social media and personalized campaigns [44]. For instance, the global fast-food chain McDonald’s ventured into mobile advertising in July 2022, launching its first campaign that integrated social media and rich media technologies. These ads were placed on popular platforms, including Facebook, Twitter, and the National Football League (NFL) mobile sites [49]. Similarly, Macy’s strengthened its mobile commerce initiatives by utilizing interactive advertising and continuously testing its online sales strategies throughout 2021. All of the mentioned potential factors of mobile ads are given in

Table 1. List of factors of mobile ads.

No.	Name of the Factors	Definition	Supports
A1	Involvement	The extent to which customers engage with mobile advertisements.	[43,44]
A2	Language	The linguistic style and wording utilized in mobile advertisements.	[43,44]
A3	Type of website	The alignment of advertised products with the website where the ads are displayed.	[43,44]
A4	Information Privacy	The safeguarding of users’ data and privacy by mobile advertisements.	[29]
A5	Entertainment	Mobile ads evoke joy, delight, and smiles, positively influencing people.	[42,45,46]
A6	Irritation	Mobile ads generate annoyance, aversion, and other undesirable reactions.	[42,44,45]
A7	Perceived Usefulness	Consumers perceive the information in mobile advertisements as useful.	[42,44,45]
A8	Perceived ease of use	Mobile advertisements are user-friendly and effortless to navigate.	[44,45]
A9	Credibility	Mobile ads create a sense of credibility and dependability for consumers.	[44,45,46]
A10	Price	Mobile ads offer comprehensive pricing details for products and services.	[22,44]
A11	Preference	Mobile ads provide personalized information tailored to users’ interests.	[22,44]
A12	Promotion	Mobile ads incorporate promotional campaigns targeted at customers.	[22,44]
A13	Interest	Mobile ads present content relevant to users’ individual preferences.	[22,44]
A14	Brand name	Mobile ads convey details regarding particular brands.	[22,44]
A15	Mobile device	Mobile ads distribute promotional messages that match the brand of the user’s mobile device.	[22]
A16	Informativeness	Mobile advertisements supply consumers with substantial and sufficient information.	[44,46]
A17	Incentives	Mobile ads share promotional details through incentives like coupons, discounts, freebies, and rewards.	[46]
A18	Social media	Mobile ads leverage social media platforms for marketing campaigns.	[48]
A19	Rich media	Mobile ads include diverse multimedia formats (e.g., animations, audio, or videos) to communicate advertising messages.	[47]
A20	Game-based	Mobile ads are embedded within games as part of marketing strategies.	[44,48]

.

3. Methods

3.1. Implemental Procedure

Feature selection plays a crucial role in data mining and machine learning, particularly in terms of handling high-dimensional data, which can negatively impact the learning performance of machine learning models [18,50]. To mitigate this, it reduces dimensionality and enhances accuracy by eliminating irrelevant or redundant features [51,52,53,54].

This study adopts SVM-RFE to identify key factors affecting user loyalty in mobile advertising. Originally used in gene selection for cancer classification [55,56], SVM-RFE applies the Support Vector Machine algorithm to assign weight rankings and iteratively removes the least significant features [52,57,58]. Widely used across fields, SVM-RFE has been applied in electricity price forecasting [59], gas sensor optimization [50], and Alzheimer’s prediction [60]. Albashish et al. [58] proposed BBO-SVM-RFE by integrating Binary Biogeography Optimization, while Chang et al. [57] used it with LASSO to identify trust criteria. It also helped identify live-streaming behavior predictors [61] and gender recognition features using gait data [62].

Tseng and Wei [47] found that integrating SVM-RFE with other methods enhances effectiveness. Thus, following Figure 1, this study applies SVM-RFE alongside two other techniques to extract critical mobile advertising factors influencing loyalty. This research builds upon previous successful studies using feature selection to handle data collected from questionnaires. For example, ref. [63] employed feature selection techniques to identify key factors influencing online learners’ performance based on data collected from a Kano-style questionnaire. Similarly, ref. [64] successfully identified critical factors of in-app purchases using feature selection methods such as SVM-RFE, DT, and IG. Ref. [65]’s IS-DT feature selection method was proposed to identify important service quality factors in programmatic buying from questionnaire data. Therefore, this study also applies feature selection methods to process data collected through questionnaires.

The implemental procedure has been clarified step by step as follows:

Step 1: Define candidate mobile ad factors

This study will collect current studies from the literature on mobile ads, collect possible factors from the literature, attempt to redefine mobile ad factors, and use these factors to design a questionnaire. The defined independent variables and their definitions are listed in Table 1. In addition, we use “repurchase intention” to measure customers’ loyalty.

Step 2: Design the questionnaire

This step began by designing a questionnaire after identifying potential mobile ad factors from the above literature on mobile ads. Each mobile ad factor was divided into two questions, found in Appendix A, with the two questions and definitions overlapping. The questionnaire is divided into two parts. Part one is about the factors that affect mobile ads and the other five scales; part two is about the basic survey of the subjects on mobile ads and the basic information of the subjects. The developed question items are provided in Appendix A.

To show how to measure independent and dependent variables, here is a brief illustrative example. We take the “game-based (No. 20)” factor as an example. According to the definition mentioned in Table 1, we can create 2 question items, which can be found in Appendix A, for measuring the independent variable and the other question item for measuring the dependent variable, as follows:

Independent variable (No. 20: game-based)

Q 39: If a mobile advertisement adopts the style of interactive mini-games (such as jigsaw puzzles, word solitaire, and guessing riddles), what do you think of it?

(A) Very important; (B) Important; (C) Neutral; (D) Unimportant; (E) Very unimportant.

Q 40: What do you think about whether the mobile advertisement is linked to the game app's content, or if the advertisement screen appears after the game ends?

(A) Very important; (B) Important; (C) Neutral; (D) Unimportant; (E) Very unimportant.

Dependent variable

Q: If an advertisement includes the above factors that you think are important or very important, and you have bought similar products (services) in this ad, will you repurchase these products (services) because you have seen this ad?

(A) Very unwilling; (B) Unwilling; (C) Neutral; (D) Willing; (E) Very willing.

Step 3: Pre-test

At this stage, the designed mobile ads questionnaire should be tested first. The questionnaires were sent to the targeted population using the Snowball sampling method in Taiwan. This method is used to distribute to the acquaintances of the subjects because the primary distribution focus is on adolescents (ages 15 to 29). Those who know them collaborate to enhance their factors. According to the respondents’ feedback, this mobile ad questionnaire was revised.

Step 4: Data collection

Data for this study were gathered both offline and online using a set of questionnaires. In order to gather the respondents’ perspectives, we circulated a set of questionnaires using the snowball sampling technique.

Step 5: Data preparation

The ordinal data of the independent variable will be encoded using values from 1 to 5 and then normalized to a range between 0 and 1. The dependent variable is repurchase intention, which is represented by 1 and 0 to indicate the presence or absence of the intention to repurchase, respectively. At this stage, the collected questionnaires are subjected to a 5-fold cross-validation method, which is divided into five equal parts, of which four parts are used as the training set, the remaining one part is used as the test set, and then five rounds are performed. The 5-fold cross-validation method can be used in the results of three feature selection methods, and the performance evaluation is more accurate.

Step 6: Feature selection

In this step, the customer loyalty variable is selected using a voting mechanism and three feature selection methods to represent the critical factors of mobile ads. The process of SVM-RFE, correlation-based, and consistency-based feature selection methods has been given in Section 3.2, Section 3.3 and Section 3.4.

Step 7: Building the SVM model

For each extracted feature subset, we will use it to build the SVM model. Suppose the smaller feature subset can perform better or have a classification similar to the original feature subset. In that case, the smaller feature subset obtained by the feature selection method has more or similar information. We know that the selected features are more important.

Step 8: Performance evaluation and drawing conclusions

This step evaluated the performance of different feature selection methods, followed by a sensitivity analysis. Finally, this study concludes with implications of the research and practice based on this evaluation.

3.2. SVM-RFE

We used the SVM-RFE method [58,66]. The SVM-RFE algorithm is given as follows in Equations (1)–(13).

Inputs:

Training examples

$$X_0={[X_1,X_2,\dots ,X_k,\dots ,X_l]}^T$$

(1)

Class labels

$$y={[y_1,y_2,\dots ,y_k,\dots ,y_l]}^T$$

(2)

Initialize:

Subset of surviving features

$$s=[1,2,\dots ,n]$$

(3)

Feature ranked list

$$r=[]$$

(4)

Repeat until

$$s=[]$$

(5)

Restrict training examples to good feature indices

$$X=X_0(:,s)$$

(6)

Train the classifier

$$\mathsf{\alpha }=SVM-train(X,y)$$

(7)

Compute the weight vector of dimension length(s)

$$W={\displaystyle \sum _k{\mathsf{\alpha }}_k{y}_k{x}_k}$$

(8)

Compute the ranking criteria

$$c_i={(w_i)}^2, for all i$$

(9)

Find the feature with the smallest ranking criterion

$$f=\arg \min (c)$$

(10)

Update the feature ranked list

$$r=[s(f),r]$$

(11)

Eliminate the feature with the smallest ranking criterion

$$s=s(1:f-1, f+1; length(s))$$

(12)

Output:

Feature ranked list

$$r$$

(13)

3.3. Correlation-Based Method

This method regards each feature as an individual with the ability to predict. When the features of the subset are related to each other, the features in the subset are highly related to the category. The correlation-based method is a straightforward filter algorithm. This method will retain the highly related features to the class, while the poorly related features will be ignored, and then the retained features will be accepted by this method. The core Equation (14) is based on the correlation method [47].

$$M_s={\displaystyle \frac{k\overline{r_{cf}}}{\sqrt{k+k(k-1)\overline{r_{ff}}}}}$$

(14)

M_s is an advantage that contains $k$ features in the feature subset S; $\overline{r_{cf}}$ is the mean of feature-class associations ($f$$\in$ S), $\overline{r_{ff}}$ and is the mean of the internal associations between features and other features; the equation can be considered a set of eigenvalues that can predict the class.

The best first method determines a greedy hill-climbing algorithm with backtracking, a regional search optimization technology that can efficiently evaluate every attribute. By adding or deleting a single attribute, its search direction can be divided into forward search and backward search. Greedy stepwise is a greedy search for the attribute subsets owned by a space. This method is similar to Best First, and it can also search in both directions, but it does not use backtracking; instead, it adds or deletes when approaching the critical point to retain the best subset of attributes. In addition, the linear forward selection is a method extended from the best first search strategy. However, it reduces the number of evaluations in the search and uses a rigorous method to generate the final subset. This method limits the expansion of the number of attributes at each stage. The method is used to determine the optimal subset size. Further, genetic search uses a simple genetic algorithm [67] to start the search by selecting attributes indexed by the number of generations, probabilities of crossover and mutation, and population size search strategy [68].

3.4. Consistency-Based Feature Selection Method

This method projects the training samples onto the subset. The attributes of the subset achieve a consistency level in the category value, and then the value contained in the attributes of the subset is evaluated. This method uses the Las Vegas Filter (LVF) algorithm [69] which is shown as below (Algorithm 1) to help the method efficiently filter out the results when selecting features. This method uses γ (allowable inconsistency ratio) to establish a standard for feature selection, reducing dimensions, and its calculus. The method is given as follows.

Algorithm 1 LVF algorithm

Enter: MAX-TRIES,          D—dataset;              N—number of attributes;              γ—allowable inconsistency rate;       Output: A set of features that satisfy the conformance criteria. Cbest = N; For i = 1 to MAX-TRIES       S = random Set(seed);       C = num of Features(S);       If (C<Cbest)              if (Incon check (S, D) < γ)                    Sbest = S; Cbest = C;                    print_Current_Best(S);       Else if (C = Cbest) and (Incon check (S, D) < γ))                    print_Current_Best(S)       end for

4. Results and Discussion

4.1. Summary of Collected Data

In total, 482 questionnaires were filled out for this study. The sample gathered for fundamental data analysis is shown in

Table 2. Basic data analysis.

Title	Scale	Percentage (%)
Gender	Male	64.11%
	Female	35.89%
Highest education	Below middle school	0.00%
	High school (vocational)	2.70%
	University/College	79.67%
	Institute and above	17.63%
Age	Under 19	28.01%
	20~29 years old	65.98%
	30~39 years old	2.70%
	40~49 years old	2.49%
	50~59 years old	0.62%
	Over 60 years old	0.21%
Average monthly income	Less than NTD 20,000	83.61%
	Between NTD 20,000 and 40,000	11.41%
	Between NTD 40,000 and 60,000	4.36%
	More than NTD 60,000	0.62%
Use a mobile device (every day)	Between 0 and 3~ h	28.84%
	Between 3 and 6~ h	45.44%
	Between 6 and 9~ h	18.46%
	More than 9 h	7.26%
Number of purchases made using mobile devices (half a year)	Within 5 times	75.93%
	5~10 times	15.77%
	10~20 times	5.81%
	More than 20 times	2.49%
Number of clicks on mobile ads (half a year)	0~3 times	60.17%
	3~10 times	30.50%
	10~30 times	5.80%
	More than 30 times	3.53%
Top reasons for recent clicks on mobile ads (check)	Advertising is creative	38.17%
	Ads are interactive	15.35%
	Advertising is useful	16.80%
	Ads attract me	43.36%
	Interested in advertised products	45.64%
	Accidentally pressed	53.53%
	Never ordered	10.79%
Where to find mobile ads (check)	App	65.35%
	Shopping site	41.08%
	Facebook	70.54%
	Twitter	5.81%
	Youtube	54.56%
	Newsletter	10.17%
	Other	2.28%
Purchase behavior	Immediate purchase	8.30%
	Consider buying in the future	80.70%
	Do not consider buying	11.00%

. Most of the sample participants are male, with a percentage of 64.11%; their highest level of education is primarily a university or junior college; and their age range is primarily 20 to 29 years old, with a percentage of 65.98%; their average monthly income is primarily less than NTD 20,000 (New Taiwan Dollars), with a percentage of 83.61%; and their average daily time spent on mobile devices is primarily 3 to 6 h. Their number of mobile device purchases in a half-year is primarily less than five. In addition, the number of clicks on mobile ads in half a year is mainly 0 to 3 times, followed by 3 to 10 times, accounting for 60.17% and 30.50%, respectively. The reasons for clicking mobile ads recently, except for accidental clicks, are that they are interested in the advertised products, and that the advertisements attracted them. Accounted for 45.64% and 43.36%, the respondents found mobile ads mainly on Facebook, followed by Apps and YouTube video and audio websites, at percentages of 70.54%, 65.35%, and 54.56%, respectively. Their consumption behavior model is primarily future-focused, with purchases making up 80.70%.

The basic data analysis results show that most mobile ad locations are app-based. In-app mobile ads are preferred by 82% of developers [11], followed by the social networking site Facebook. It has been discovered that Facebook has a high exposure rate for mobile ads, allowing mobile device users to view them. It is simple to locate mobile ads. The third option is the video-sharing website YouTube. According to Susan Wojcicki, Senior Vice President of Google Advertising, nearly 50% of mobile device (mobile and tablet) users currently use YouTube (of which 50% are active) as YouTube is a repository for many mobile ads [15]. Therefore, we focus on YouTube platform. Moreover, mobile ads are often clicked unintentionally, as noted by the product manager of mobile display ads at Google, despite the increasing size of mobile phone and tablet screens. It is still easy to accidentally slide on or click links or advertisements one does not want to see. According to third-party research data, up to 50% of mobile ad clicks are accidental, which is not much different from the results of this study [14].

As a result, Google has implemented various measures to prevent false clicks, leading to a decrease in the click rate of floating reports and increased costs for advertisers. For example, in mobile image advertisements, Google identifies the picture frame as a high misclick area because users who click on adjacent content or are scrolling may accidentally touch it. It is recommended that users click on the center of the image to browse the advertising website or application, as this can reduce the burden on mobile device users and help report their click-through rate more accurately [12]. The number of samples collected in this study was tested for reliability and validity, as shown in

Table 3. Reliability analysis.

Factors	Cronbach’s Alpha	Factors	Cronbach’s Alpha
Involvement	0.887	Preference	0.880
Language	0.885	Promotion	0.880
Type of Website	0.887	Interest	0.880
Information Privacy	0.893	Brand Name	0.884
Entertainment	0.884	Mobile Device	0.882
Irritation	0.889	Informativeness	0.880
Perceived Usefulness	0.882	Incentives	0.881
Perceived ease of use	0.879	Social Media	0.884
Credibility	0.883	Rich Media	0.882
Price	0.881	Game-based	0.888

. KMO and Bartlett were used to test their validity, and the value was 0.902 *** (when its p value was less than 0.001, it was three stars). Cronbach’s Alpha value indicates that the data can correctly measure the characteristics that this research requires. In addition, Cronbach’s Alpha value is used to test its reliability, and the values of its factors are all greater than 0.7 [20], which indicates good reliability and that the data are stable and consistent.

4.2. Results of Feature Selection

In this step, we used SVM-RFE with the rankers’ opinions. The ranker calculated each input factor, generated a Merit value, and arranged it in order from the best to the worst. This research also uses this ranking to select the top 10 positions identified by each fold as important factors. The factors that appear in all five folds using the voting mechanism are retained and become important factors of the mobile ads screened out by this method.

Figure 2 summarizes the results of the SVM-RFE selection method using Equations (1)–(13). Each fold uses a ranker. After finding the important factors, the ranker is used to sort the factors from the best to the worst according to the generated merit value, and a voting mechanism is performed. The factors are found in each fold. All factors appear to keep the selected mobile ads as the important group for this method, with a 100% probability. The important mobile ads factors affecting the click rate of this method are customer loyalty factors numbered 3, 6, 11, 18, 20, 22, 24, and 38.

Secondly, we used a correlation-based method to select the important factors affecting mobile ad loyalty. This method uses four search methods: best first, genetic search, greedy stepwise, and linear forward selection, as discussed in Section 3.2. A search strategy combines this approach and a voting mechanism to filter out important factors.

Table 4. Selection factors based on the correlation method.

Methods	Fold1	Fold2	Fold3	Fold4	Fold5
Best First	3,12,14,15, 16,18,19,20, 22,23,25,30, 31,34	3,13,15,16, 18,20,22,25, 30,34	3,15,16,18, 19,20,22,23, 25,30,34,37	3,13,14,15, 16,18,20,21, 22,25,30,31, 32,34,38	3,12,14,15, 16,18,19,20, 21,22,23,25, 30,34,37
Genetic Search	3,10,15,16, 18,19,21,22, 23,34,37	15,16,18,19, 20,21,22,25, 26,31,34	3,15,16,18, 19,20,21,22, 23,26,34,37	12,14,16,18, 19,20,21,22, 25,30,34,37	15,16,18,19, 20,22,26,34
Greedy Stepwise	3,12,14,15, 16,18,19,20, 22,23,25,30, 31,34	3,13,15,16, 18,20,22,25, 31,34	3,15,16,18, 19,20,22,23, 25,30,34,37	3,13,14,15, 16,18,20,21, 22,25,30,31, 32,34	3,12,14,15, 16,18,19,20, 21,22,23,25, 30,34,37
Linear Forward Selection	3,12,14,15, 16,18,19,20, 22,23,25,30, 31,34	3,13,15,16, 18,20,22,25, 30,34	3,15,16,18, 19,20,22,23, 25,30,34,37	3,13,14,15, 16,18,20,21, 22,25,30,31, 32,34,38	3,12,14,15, 16,18,19,20, 21,22,23,25, 30,34,37
voting mechanism	3,15,16,18, 19,22,23,34	15,16,18,20, 22,25,34	3,15,16,18, 19,20,22,23, 34,37	14,18,20,21, 22,25,30,34	15,16,18,19, 20,22,34
	15 (80%), 16 (80%), 18 (100%), 19 (60%), 20 (80%), 22 (100%), 34 (100%)

Note: Four search methods are used for each fold, and the repeated factors will be filled in. Finally, the factors with more than three repetitions in 5Fold experiments are the important factors selected by this method.

is the analysis results of the correlation method using Equation (14). Each fold adopts four search strategies. After each important factor is found, the first voting mechanism is carried out, and the important factors represented by each fold are screened out. The second voting mechanism is repeated more than three times (with a probability of more than 60%), and the matching factors are unified, which is the final selection of the important factor group of mobile ads for this method. This method is important in affecting customer loyalty. However, according to Figure 3, the mobile ads factors are customer loyalty are 15, 16, 18, 19, 20, 22, and 34.

Finally, we used a consistency-based method to select three important factors that affect mobile ads, depending on the variables. This method uses four search methods: best first, genetic search, greedy stepwise, and linear forward selection. The search strategy incorporates this approach and utilizes a voting mechanism to filter out important factors.

The results of the consistency method analysis are shown in

Table 5. Selection factors based on the consistency method.

Methods	Fold1	Fold2	Fold3	Fold4	Fold5
Best First	3,9,13,16,19, 20,21,22,23, 25,30,31,38	3,9,13,14, 17,19,20,21, 22,23,25,30, 31,32,34,38	3,13,14,16, 17,19,20,21, 22,23,31,34	3,9,14,17,20, 22,23,25,30, 31,37,38	3,13,14,15, 17,20,21,22, 25,30,31,38
Genetic Search	3,5,6,9,17,19, 20,21,23,25, 30,31,37,38	2,3,5,12,13, 14,15,19,21, 22,23,25,27, 30,31,35,38, 39	3,9,12,14,16, 19,21,22,25, 29,31,32,37, 38	3,9,10,13,14, 20,23,25,27, 30,31,35,38	3,4,9,10,14, 15,19,18,19, 20,21,23,30, 31,35,37,38
Greedy Stepwise	3,9,13,16,19, 20,21,22,25, 30,31,38	9,13,14,15, 19,20,21,22, 23,25,30,31, 32,34	3,10,13,14, 16,17,19,20, 21,22,23,31, 34	3,9,14,17,20, 22,23,25,30, 31,37,38	3,13,14,15, 17,20,21,22, 25,30,31,38
Linear Forward Selection	9,13,19,20, 21,22,23,25, 30,31,32,34, 38	3,9,13,14, 17,19,20,21, 22,23,25,26, 27,30,31,34, 38	3,9,13,14,16, 17,20,21,22, 23,25,31,34, 38	3,14,16,20, 21,22,23,25, 30,31,34,37, 38	3,12,14,15, 18,19,20,21, 22,23,31
voting mechanism	9,19,20,21, 25,30,31,38	13,14,19,21, 22,23,25,30, 31	3,14,16,21, 22,31	3,14,20,23, 25,30,31,38	3,14,15,20, 21,31
	3 (60%), 14 (80%), 20 (60%), 21 (80%), 25 (60%), 30 (60%), 31 (100%)

Note: Four search methods are used for each fold, and the repeated factors will be filled in. Finally, the factors with more than three repetitions in 5Fold are the important factors selected by this method.

. For each fold, four search strategies are used. After discovering the important factor, the first voting mechanism is used, and the important factors represented by each fold are screened out. In Figure 4, we carry out the second voting mechanism, repeat it more than three times (that is, more than 60% probability), and integrate the matching factors, that is, the final selection of important factors for mobile ads for this method is represented by the numbers 3, 14, 20, 21, 25, 30, and 31.

4.3. Performance Evaluation

Table 6. Summary of selected factors by three feature selection methods.

Factors	No.	Consistency	Correlation	SVM- RFE
Involvement	1
	2
Language	3		V	V
	4
Type of Website	5
	6			V
Information Privacy	7
	8
Entertainment	9
	10
Irritation	11			V
	12
Perceived Usefulness	13
	14		V
Perceived ease of use	15	V
	16	V
Credibility	17
	18	V		V
Price	19	V
	20	V	V	V
Preference	21		V
	22	V		V
Promotion	23
	24			V
Interest	25		V
	26
Brand Name	27
	28
Mobile Device	29
	30		V
Informativeness	31		V
	32
Incentives	33
	34	V
Social Media	35
	36
Rich Media	37
	38			V
Game-based	39
	40

summarizes the results of three feature selection methods. We use two variables to represent each factor and three methods to filter the variables. Finally, the selected variables point to the original variables. The factors that have been selected by more than one method are “language”, “credibility”, “price”, and “preference”. Figure 5 provides a simple Venn diagram to show overlapping key features across feature selection methods

Next, we use an SVM classifier to evaluate the performance of selected feature subsets.

Table 7. Performance evaluation by SVM.

Original Feature Set (Without Implementing Feature Selection)—40 Variables
Indicators	Fold1	Fold2	Fold3	Fold4	Fold5	Avg.(std.)
Accuracy (%)	62.50	67.71	70.83	69.79	69.79	68.12(3.34)
Precision (%)	61.80	68.50	69.40	72.60	69.20	68.30(3.96)
Recall (%)	62.50	67.70	70.80	69.80	69.80	68.12(3.34)
F1 (%)	60.80	64.00	66.90	67.00	68.10	65.36(2.97)
Time (s)	13.68	32.10	13.70	25.35	23.51	21.67
Consistency—7 variables
Accuracy (%)	62.50	66.67	67.71	69.79	67.71	66.88(2.70)
Precision (%)	61.80	66.67	67.00	77.40	66.80	67.93(5.72)
Recall (%)	62.50	66.67	67.70	69.80	67.70	66.87(2.70)
F1 (%)	60.80	63.20	67.30	65.40	66.60	64.66(2.66)
Time (s)	5.04	8.16	8.33	8.78	9.70	8.00
Correlation—7 variables
Accuracy (%)	63.54	66.67	72.92	70.83	70.83	68.96(3.78)
Precision (%)	63.30	66.70	71.80	74.70	70.20	69.34(4.44)
Recall (%)	63.50	66.70	72.90	70.80	70.08	68.80(3.71)
F1 (%)	61.20	63.20	71.80	67.90	69.70	66.76(4.44)
Time (s)	10.03	4.88	8.97	8.08	9.11	8.21
SVM-RFE—8 variables
Accuracy (%)	60.42	65.63	65.63	68.76	70.83	66.25(3.94)
Precision (%)	59.40	64.80	64.40	72.90	70.60	66.42(5.37)
Recall (%)	60.40	65.60	65.60	68.80	70.80	66.24(3.95)
F1 (%)	58.60	62.90	64.80	65.10	69.00	64.08(3.78)
Time (s)	5.27	7.69	8.53	9.36	7.89	7.75

lists the classification performance of SVM using the original feature set without implementing feature selection and the results of using three feature subsets using three feature selection methods. It could be considered as a comparison base. From Table 7, in the original feature set, we find that the mean values of accuracy and F1 are 68.12% and 65.36%, respectively. Moreover, the training time is 21.67 s on average.

In the consistency-based method, it can be found that Fold4 has the highest classification accuracy, and Fold1 takes the least time. The average accuracy is slightly lower than the original feature set, and F1 is slightly improved. However, the average training time is reduced to 8 s from 21.67 s. In the correlation-based method, the performance improved, regardless of accuracy and F1. Furthermore, the average training remains at 8.21 s. In SVM-RFE, the average training time reaches 7.75 s. Both accuracy and F1 are slightly lower than the original feature selection. The comparison of training time can be found in Figure 6.

Regarding efficiency, t-tests on training time at the 95% confidence level have been performed as shown in

Table 8. Hypothesis testing for efficiency.

Hypotheses	p-Value	Decision
$\begin{array}{c}{H}_0: {\mu }_{full}\ge {\mu }_{consistency}\\ H_1: {\mu }_{full}>{\mu }_{consistency}\end{array}$	0.01	Reject H0
$\begin{array}{c}{H}_0: {\mu }_{full}\ge {\mu }_{correlation}\\ H_1: {\mu }_{full}>{\mu }_{correlation}\end{array}$	0.011	Reject H0
$\begin{array}{c}{H}_0: {\mu }_{full}\ge {\mu }_{SVM-RFE}\\ H_1: {\mu }_{full}>{\mu }_{SVM-RFE}\end{array}$	0.009	Reject H0

. The p-values for the three groups were 0.01, 0.011, and 0.009, respectively, all below 0.05. These results confirm that applying dimensionality reduction via consistency, correlation, or SVM-RFE significantly reduces model training time, demonstrating the efficiency of the reduced models.

The original feature set contains 40 variables, which are reduced to 8, 7, and 7 after feature selection. However, regardless of accuracy or F1, its performance gap is insignificant. Among them, the results of the correlation-based method are even better, which means that the same or even better classification performance can be achieved using fewer variables. That means the selected features are important.

4.4. Sensitivity Analysis

In addition to using performance evaluation to screen important factors for the three feature selection methods, sensitivity analysis is also used to verify performance, which uses the opposite factor group (those factors are not selected by the feature selection method) to classify. To compare whether the performance is higher than the performance of the selected feature subset by feature selection methods. The feature selection method is unsuitable if its performance exceeds the selected feature subsets.

Sensitivity analysis will take longer than the important factor groups selected by the three feature selection methods because we used all features except the selected feature subsets. The results are shown in

Table 9. Results of the sensitivity analysis.

Used Features Indicators	Selected Factors	Unselected Factors	Pass or Not
Consistency
Accuracy (%)	66.88(2.70)	68.34(2.51)	Not
Time (s)	8.00	18.90
Number of used features	7	33
Correlation
Accuracy (%)	68.96(3.78)	67.50(2.59)	Pass
Time (s)	8.21	21.06
Number of used features	7	33
SVM-RFE
Accuracy (%)	66.25(3.94)	66.46(1.36)	Pass
Time (s)	7.75	20.40
Number of used features	8	32

. As can be seen from the table below, the correlation-based method and SVM-RFE have passed the sensitivity analysis. The correlation-based method and SVM-RFE are suitably used as feature selection tools in this case.

At a 95% confidence level, t-tests were conducted to compare classification model accuracy and training time using the original features versus only the selected important features across three methods: consistency, correlation, and SVM-RFE. The results showed that the p-values were all greater than 0.05, indicating no significant difference in classification performance between models built with reduced features and those using all. This suggests that the selected subset of features retains most of the information in the complete set, making them important features. Moreover, for all three methods, the models built with fewer important features required significantly less training time than models using the complete feature set.

5. Conclusions

The present study tried to identify potential factors of mobile ads. Then, we applied SVM-RFE, correlation-based, and consistency-based feature selection methods to identify the critical factors that directly improve customer loyalty. Our results indicated that the correlation-based and SVM-RFE methods outperform the consistency-based feature selection methods. The correlation-based method selected seven critical factors that can improve customer loyalty, namely “language”, “perceived usefulness”, “price”, “preference”, “interest”, “mobile device”, and “informativeness”. Among them, price and preference were selected by all feature selection methods.

Table 10. The selected important factors and the given suggestions.

Selected Factors	Suggestions	Selected Factors	Suggestions
Price	Mobile ads should give customers clear product price information, such as displaying the product’s price tag in the mobile advertisement.	Interest	Mobile ads should show products or services that customers are interested in. For example, mobile ads can provide advertisements for video games and beauty products according to gender and launch advertisements for their favorite products according to personal preferences.
Preference	Mobile ads should provide customers with their preferred products or services, as mobile ads can distribute possible advertisements according to the customer’s location (car parking area).
Language	Mobile ads can increase customer loyalty if they use text-based language that is friendly to viewers, such as local catchphrases or language that incorporates current events.	Mobile device	Mobile ads should be served with customer-owned devices, e.g., ads for products of the same brand on customer-owned branded devices.
Perceived usefulness	Mobile ads should be helpful to customers, such as providing customers with the products they need to search for on mobile ads or recommending products when customers need them.	Informativeness	Mobile ads should give customers enough product information. For example, mobile ads should let customers know the content of the product.

summarizes the suggestions provided.

The contributions of this study to customer loyalty research are as follows: First, the traditional conceptualization of customer loyalty was successfully applied in the new mobile commerce context. Second, the results indicated that customer loyalty is affected by “language”, “perceived usefulness”, “price”, “preference”, “interest”, “mobile device”, and “informativeness”.

This study proposed feature selection-based methods to identify the important factors that enhance customer loyalty. Compared to conventional statistical analysis approaches, feature selection can ignore the assumptions of the data used, such as independence, normality, and constant variance in regression. In addition, feature selection methods have also been widely used to analyze big data in recent years. Moreover, the feature selection method can handle various data, including text. This can make up for the shortcomings of traditional statistical methods.

5.1. Implications of the Research and Practice

This study validated existing literature within the context of customer loyalty in mobile ads and identified the critical factors to improve customer loyalty through mobile ads using three feature selection methods, including SVM-RFE, correlation-based, and consistency-based methods, for the first time. Hence, our implications of the research are as follows: Our work is significant as the online world has become more prevalent in customers’ lives in the past decade and even more so during the COVID-19 pandemic. Our contribution is threefold. First, as far as the author knows, this is the first study conducted in the context of improving mobile ads to improve customer loyalty. We used novel datasets to estimate customers’ preferences for mobile ads, and our study used a larger sample representative of the present generation (who frequently use mobile apps for their shopping). Second, we explored the important factors of mobile ads using three techniques. We considered how proposed user targeting methods affect customers’ interests, and how they are helpful to advertisers and AdTech companies whose services make up the online advertising ecosystem. Third, feature selection methods based on SVM-RFE, correlation-based, and consistency-based feature selection methods were presented in this study. As a result, the presented methodology differs from previous studies’ methodologies and adds a new method to the mobile marketing literature. From a managerial and policy perspective, our research speaks to the future of the programmatic ad sector. It indicates the best prediction of what customers want from the other stakeholders in the online market. Finally, the findings of this study provide the most important factors for increasing customer loyalty to advertisers and promoting agencies.

Several practical implications for advertisers and promoting agencies are presented, and various important factors across the mobile ads process are identified. First, based on our results, “Price” and “Preference” were identified as the most important factors in improving customer loyalty. Hence, management attention may be more fruitful if focused on such factors in their advertising processes. Accordingly, to increase customer loyalty, advertisers and managers need to be concerned about the experience of consumers, from the first encounter through purchasing to delivery and beyond, as this can influence consumer satisfaction, thus influencing customer loyalty. Second, prior research has shown that information quality, website type, and media may influence customer loyalty and improve sales through mobile ads. This demonstrates the significance of value as a strategic goal from a managerial perspective. The importance of employing correlation-based SVM-RFE and this study’s empirical results further highlight consistency-based feature selection methods. We replaced the traditional qualitative analysis that used questionnaires with feature selection approaches. This approach’s outcome promotes client loyalty. Customer loyalty should be included in their current loyalty valuation techniques, and corrective action should be taken to improve it.

5.2. Practical Implications

The findings of this study offer several practical implications for advertisers, mobile app developers, and digital marketing strategists. By identifying “Price” and “Preference” as the most critical factors influencing customer loyalty across all three feature selection methods, businesses are encouraged to tailor mobile ads that provide transparent pricing and align with consumer preferences. For example, leveraging location-based services or browsing histories to recommend relevant products can significantly enhance ad relevance and customer engagement. A concrete tactic might involve sending a push notification offering a 10% discount on items left in a user’s cart, triggered when the user is near a physical store location or at a specific time of day.

Additionally, other consistently influential factors—such as “Language”, “Perceived Usefulness”, “Interest”, “Mobile Device”, and “Informativeness”—highlight the importance of content quality and personalization. Advertisements using localized language, clear value propositions, and device-optimized formats can foster stronger user connections and reduce ad fatigue. Mobile ads should go beyond mere exposure to provide genuinely helpful content that supports customers in their decision-making process.

Furthermore, integrating feature selection techniques into mobile marketing analytics demonstrates a promising direction for optimizing ad performance. Businesses can adopt these data-driven methods to refine large sets of user data and isolate key drivers of consumer behavior without relying on assumptions typical in traditional statistical models. This approach improves marketing efficiency and supports strategic investment in content development.

Lastly, given the high click rates on platforms like Facebook, YouTube, and in-app environments—as observed in the data analysis—marketers should prioritize these channels for campaign deployment. Ensuring that mobile ads are interactive and minimally intrusive can help address the common issue of accidental clicks, enhancing the accuracy of engagement metrics and overall campaign effectiveness.

5.3. Limitations and Future Direction

This study presents several limitations that warrant consideration. First, while it is among the first to apply SVM-RFE, correlation-based, and consistency-based feature selection methods to identify key factors affecting customer loyalty in mobile advertising, the findings are derived from a single empirical study. Generalizing these results to other ad formats or user groups—such as offline ads or older consumers—should be performed cautiously. Moreover, firm-level variables such as relative price, market share, and profitability, which are known to correlate with customer loyalty, were not included in the analysis.

Second, the sample is highly skewed toward young Taiwanese participants, with most respondents aged 15 to 29 and primarily students. Although this demographic aligns with mobile and online gaming audiences, it limits the broader applicability of the findings across diverse populations. To mitigate potential response bias, subgroup-level factor loading assessments were conducted to evaluate variance homogeneity and detect latent response bias. However, the snowball sampling technique may also introduce bias, as it generates homogeneous participant pools with similar traits and preferences.

Future research should adopt more rigorous sampling methods, such as stratified or quota sampling, to enhance generalizability and robustness and target a broader demographic across regions and industries. Cross-cultural studies that include more diverse variables can also provide deeper insights into the dynamics of mobile ad effectiveness and customer loyalty. In addition, future studies could also use model-agnostic interpretability techniques (e.g., SHAP values [69]) to serve as a reference model for validating feature importance beyond SVM-RFE or correlation filters.

In future studies, the researcher can use various feature selection methods or big data tools to improve customer loyalty. Additionally, we only used 482 samples in this study. Future works can increase the data size to sample representative examples.