A Fuzzy Rule-Based Framework for Consumer Behavior Analytics

Abstract

Understanding consumer behavior is vital for businesses seeking to personalize services, optimize marketing strategies, and improve customer retention. However, analyzing such behavior at scale presents significant challenges due to the volume, velocity, and variety of data, as well as the need for accurate and interpretable prediction models. Traditional classification methods often fall short when applied to large-scale, high-dimensional behavioral datasets, leading to issues in scalability, accuracy, and real-time processing. To address these limitations, this paper introduces a novel framework for consumer behavior analysis using an Improved Fuzzy Classification with Bagging and MapReduce Coordination (IFCBMC) approach, specifically designed for big data environments. The primary objectives of this research are: (1) to develop a scalable classification model suitable for distributed data processing, (2) to enhance prediction accuracy through fuzzy rule-based learning, and (3) to evaluate the robustness of the proposed model against existing state-of-the-art classifiers. The process begins with data preprocessing, including cleaning and modified normalization, followed by distribution of the data across a MapReduce architecture to manage scale and speed. Extracted features from multiple data partitions (mappers) are aggregated and processed by an enhanced fuzzy rule-based classification model. To improve prediction robustness, a bagging ensemble strategy is applied, where multiple classifiers are trained on different data subsets, and the best-performing models are randomly selected and merged during the reduce phase. The proposed IFCBMC method outperforms all compared models, achieving the highest accuracy of 0.960, significantly surpassing traditional approaches such as LSTM (0.862), LINKNET (0.866), SQUEEZENET (0.858), SVM (0.863), DCNN (0.860), Bi-GRU (0.854), RNN (0.861), and DNN (0.860).

Keywords

modified normalization pre-processing MapReduce framework sensitivity data cleaning

1. Introduction

Recent developments in big data technologies and the spread of digital platforms have completely changed how companies anticipate and comprehend consumer behavior. Large-scale datasets might potentially yield valuable insights through the use of big data technologies, but putting these technologies to use needs advanced data processing and analytical methods. An individual or group that is not actively involved in commercial or company activities but wants to order or use purchased items, goods, or products mainly for their own social, relatives, or domestic purposes is referred to as a consumer (Anand et al., 2021; Dengler & Prüfer, 2021). Focusing on every consumer's needs and demands separately is crucial because consumers are always significant in an entrepreneur's or a company's activity. This serves as an example of the idea for analyzing consumer behavior (Li et al., 2023; Reyes-Menendez et al., 2022).

An organization can learn how consumers choose a product or service by conducting a consumer behavior study (Simona-Vasilica et al., 2021). They require a combination of both quantitative and qualitative information from surveys of consumers, customer interviews, and data acquired from observation of their activity in-store and online (Hung & Van-Nam, 2023; Shiu-Li & Yi-Hsien, 2022). Due to the lack of personal contact, there is doubt regarding the product's quality when an individual buys it via a digital retailer (Abedin et al., 2023). Unexperienced buyers can cut down on their search expenses and uncertainty about the quality of the products they purchase by utilizing the knowledge from past review data (Ulph et al., 2023). Organizations can also use user-generated feedback to create, build and market new goods as well as to gauge specific customer needs, wants, fulfilment, and concerns (Sundararaj & Rejeesh, 2021).

Organizations may target certain demographics with their advertising efforts, increase customer loyalty, and spot new trends by studying customer behavior. Additionally, by using this information, organizations may stay one step ahead of the competition and adjust to shifting consumer preferences. A wide range of industries, including electronic commerce (Li et al., 2020; Mou & Benyoucef, 2021; Shiu-Li & Yi-Hsien, 2022; Tassell & Aurisicchio, 2023), marketing (Raj Kannan et al., 2021), social networking sites (Mikalef et al., 2023; Sharifi & Shokouhyar, 2021; Sundararaj & Rejeesh, 2021), tourism (Li et al., 2022), business (Tao & Zhou, 2023), financing (Abedin et al., 2023), utilities (Simona-Vasilica et al., 2021), etc., are heavily utilizing the development of consumer behavior analysis.

Decisions are mostly based on the benefits of consumer behavior analysis models, which are essentially one-dimensional (Taghikhah et al., 2021). However, those models deal only with a few limitations in some circumstances, and they assume that individuals remain the same while ignoring variances in customers and goods, values or passions, age, or gender (Xue et al., 2022; Zaremohzzabieh et al., 2021). Consumer behavior studies are important for businesses because they help them understand their target market, identify the needs of their customers, and develop effective marketing strategies that influence their decisions to buy (Bilal et al., 2020; Stangherlin et al., 2023). Previous research frequently concentrates on either conventional statistical techniques or basic big data framework applications, failing to fully utilize sophisticated machine learning models. Furthermore, in the field of customer behavior research, not many studies have thoroughly examined the combination of MapReduce for distributed processing and fuzzy rule-based models for categorization. This study is motivated by the necessity to fill these gaps by putting forward a novel IFCBMC strategy. This project aims to improve the scalability and accuracy of customer behavior prediction models by utilizing the scalability of the MapReduce framework and integrating fuzzy logic with ensemble learning through bagging-based classifiers. In this work, an IFCBMC-based consumer behavior analysis model is introduced, and the major contribution is summarized below:

Proposing the modified Normalization in the preprocessing phase. Here, the min-max normalization is immersed into the Tanh normalization process to enhance the input data by removing the redundant data and also ensuring the data are ready for further processing.

Proposing a MapReduce framework to handle big data, in which features like improved entropy and correlation-based features are extracted in the mapper phase, which enhances the uncertainty measures in the big data analysis. Particularly, the improved entropy is the combination of Modified Deng entropy and Belief Entropy.

Proposing an Improved fuzzy rule-based classification model that includes the data Bagging and Mapping criteria.

1.1 Research Objectives

To investigate the preprocessing phase of big data, focusing on data cleaning and modified normalization techniques.

To utilize the MapReduce framework to handle and process large volumes of data efficiently.

To extract and combine features from distributed data sources to enhance classification accuracy.

1.2 Research Questions

How effective is the IFCBMC approach in analyzing customer behavior compared to traditional methods?

What are the impacts of data cleaning and modified normalization on the quality of big data used for customer behavior analysis?

How does the MapReduce framework contribute to managing and processing big data in the context of customer behavior analysis?

The paper is formatted as: Section 2 represents the related works and the problem statement based on the existing models. Section 3 gives the proposed work for predicting consumer behavior using an improved fuzzy ruler-based classification model. Next, Section 4 represents the investigational outputs and analysis. Lastly, Section 5 includes the conclusion.

2. Literature Review

In 2019, Mirashk et al. (2019) has introduced a novel approach that consists of the use of big data tools and techniques based on networked computing-based methodologies. The models used to forecast consumer behavior have mostly been successful with small sets of information and parameters. In this study, RNNs were used to examine transactional data to create a model for forecasting POS user activity. RNNs were capable of forecasting high-dimensional information or time-series information. The created model was validated using actual data from one of Iran's private banks and the projected outcomes were around 87% accurate. The outcome is contrasted with earlier methods, which perform better than later-discussed methods.

In 2020, Raj Kannan et al. (2021) have developed a model for consumer behavior evaluation by using the mouse movement pattern as the basis. This aided in gathering data or mining that aided in forecasting client behavior in the online marketplace. Usually, classification methods are termed Multi-layer NN approaches and one of the efficient algorithms for data mining, named the decision tree algorithm, was used for the behavioral evaluation. These methodologies enabled precise analysis and determination of customer behavior. Several standard data sets were employed for testing and assessment, and the findings demonstrated that the suggested model provided superior analysis than previous research.

In 2020, Argyris et al. (2020) has suggested a new conceptual framework of VCSI through the Similarity-Attraction Model that was contextualized in the Social Influence literature. They identified, using VCSI, in what way Influencers create close relationships with followers by using visual congruence as illustrations of common interests in a certain field. The favorable impacts of visual consistency on followers’ brand engagement were stimulated by this implied affinity. To automatically group each image, DL techniques were then deployed. Social networking analytics were then used to reveal any undiscovered relationships between visual components and brand interaction. These outcomes from tests supported VCSI with evidence, furthering theories in the quickly developing area of multimodal information.

In 2020, Tao and Zhou (2023) has explored the forecast of business closure using digital customer evaluations and validated the models for forecasting in the food service industry, which saw an unusually high rate of attrition. The proposed method for predicting business closure included several new elements, such as the application of time-series analysis and DL methods, the use of a unique triple word embedding model for text representation, and the collection of data from online reviews through hybrid classification methods. The suggested method's examination utilizing Yelp online reviews showed improved performance in predicting business closure, demonstrating that online reviews were a reliable source of information for this purpose.

In 2021, Chaudhary et al. (2021) used the concept of big data technology to process and analyze data to forecast customer behavior on social media platforms. According to a few metrics and requirements, they examined customer behavior on social media platforms. The authors investigated consumer attitudes and perceptions of the social media network. To get high-quality outcomes, they employed a number of data preparation strategies to find mistakes, noise, outliers, and duplicate entries. To forecast consumer behavior on the social media platform, they created computational modelling applying ML.

In 2021, Jeong (2021) has examined customers’ DFs for thermostats that were programmable to recognize and forecast unexpected customer preferences, using a collection of data of 141 million Amazon assessments. This study proposed novel methods for extracting PCDs from review text information, predicting individual customer satisfaction emotions before deciding on a purchase, predicting consumers’ sentiment before they write a review, and classifying consumers’ sentiment toward a particular PCD using dependent on context word incorporation and DL models. In conclusion, the method proposed in this paper was practical, scalable and comprehensible for identifying significant factors influencing customer evaluations of various products in a specific sector and could be applied by the business to construct consumer-focused marketing approaches.

In 2022, Yuan (2022) has categorized consumer purchasing behavior features research into three distinct groups such as data-driven, theory-driven, and experience-driven. The concept of merging customer consumption behavior characteristics like happiness and devotion was offered, along with an analysis algorithm based on consumer consumption behavior. A comparative analysis showed that the data-driven strategy was the most successful in analyzing the characteristics of online customers’ purchasing patterns. Different service plans for clients with various levels of evaluation were accomplished using the choice support information base. The genetic method was used for optimizing samples to increase the sample classification accuracy and enhance the output function. The classification of samples of customer transaction data was proposed using a deep neural network structure method. It was advantageous for businesses to understand the concept of personalization as well as evaluate consumer behavior in terms of consumption and production planning.

In 2022, Li et al. (2022) has investigated the use of an ML technique to utilize social media analysis of big data for predicting interest in tourism. To anticipate the number of tourists arriving, they showed ways to gather the key themes addressed on Twitter and compute the mean emotion score for each issue as an estimate of people's overall feelings about those topics. The analysis identified important social media discussions that could be utilized to forecast Sydney visitor numbers. The analysis produced both theoretical and practical ramifications for marketing destinations and tourism research.

In 2022, Meena and Kumar (2022) have studied the efficiency of OFD organizations and the expectations of customers during the COVID-19 epidemic by using social media information. The findings, customers from India were more involved with societal responsibility than those in the US were with money-related issues. Compared to American customers, Indian consumers were generally happier about OFD companies throughout the COVID-19 pandemic. Additionally, they discovered that variables like the identity of the OFD enterprises, the scale of the market, the nation, and the COVID-19 waves were significant in modulating consumer attitudes. The study's findings provided some managerial insights.

In 2023, Abedin et al. (2023) used a variety of feature conversion approaches to translate behavioral information into different information structures, while analyzing consumer behavior and activity in the banking industry. The research that employed a real bank customer dataset consisting of 24,000 active and inactive customers offered a fresh viewpoint on the role of attribute engineering in the bank classification of clients. This paper presents a thorough, methodical analysis of the modeling of bank client behavior that could help financial service providers take the required steps to increase customer activity.

In 2024, Sitharamulu et al. (2024) have suggested a hybrid classifier model, CSDHAP, that uses the MapReduce framework to hybridize the algorithms for deer hunting optimization (DHO) and sunflower optimization (SFO) with an adaptive pollination rate. Classifiers are used in the CSDHAP, a technique for classifying data. Several criteria are used to compare the effectiveness of the suggested approach to those that are currently in use. It is noteworthy that every traditional model is surpassed by the recommended system.

In 2024, Navin and Krishnan (2024) have provided a novel mapping tool model intended to evaluate electronic health records and offer evidence-based decision support to healthcare providers. The work focuses on indicators taken from normal health exams and analyzes health information from hospital databases. The incorporation of a fuzzy rule-based classifier system into the suggested system is the fundamental component of this strategy. The existing methods of features and challenges are tabulated in Table 1.

Table 1.
Features and Limitations of Reviewed Works.

Author [Citation] Methodologies Features Challenges

Mirashk et al. (2019) RNNs to create a model for forecasting POS user activity The created model was validated using actual data from one of Iran's private banks, and the projected outcomes were around 87% accurate Need to recognize the difference between the predicted and natural behavior of customers

Raj Kannan et al. (2021) CBA-MMP To efficiently assess the patterns of consumer behavior, the suggested MMP-CBA combines the strengths of the DT algorithm with MLNN. However, its analytical capabilities appeared to be weaker because it used ML technology to operate.

Argyris et al. (2020) Visual congruence: DL algorithms for automatic image classification. The study not only used DL algorithms for image classification, but it also suggested an ensemble method as a reliable strategy to support future academics in their efforts to delve into the world of visual content that is shared on social media. It exclusively examined Instagram to show how aesthetic consistency increases consumer interaction with brands.

Tao and Zhou (2023) DL-based time–series analysis. The CNN-LSTM model had an accuracy of above 81%, making it useful for practical business closure prediction. Since its conclusions were drawn on the analysis of data from a specific industry, the model was inapplicable to other industries.

Chaudhary et al. (2021) ML-based social media consumer behavior model. The installation of social media marketing and performance analysis was combined, and the data analyses were performed using a subpar instrument. To operate on a real-time consumer basis, it was not implemented. This model's performance on real-time data would be very subpar.

Jeong (2021) ML-based consumer preference identification model. The proposed approach could be used by industry to create client-centered marketing strategies because it is adaptable, scalable and comprehensible for identifying significant drivers of consumer evaluations for various products in a particular industry. If it could distinguish between real and fake reviews posted on various internet sites, it might provide a more accurate analysis of the data.

Yuan (2022) Merging customer consumption behavior characteristics- algorithm The study mines a huge quantity of consumer consumption data for a variety of forms of useful information, including customer preferences and consumption patterns Have to utilize and larger number of samples to prove the efficiency of the algorithms

Li et al. (2022) Tourism Demand Analysis Framework The topic models and sentiment analysis methods suggested in this paper were used to extract sentiments, and the result was a social listening tool for predicting and managing tourism demand. Since the study was restricted to a specific location and relied on data from just one social media platform, Twitter, its conclusions regarding the elements that influence demand needed to be more broadly generalized.

Meena and Kumar (2022) LDA-based framework – text analysis, topic extraction, and sentiment analysis. It normally looked into subjects including social responsibility, delivery procedures, perceived consumer responsibility and COVID-19 financial impacts and it offered better-studied results. It was impossible to ascertain how people's sentiments related to the performance of OFD enterprises on the stock market throughout the COVID-19 period.

Abedin et al. (2023) Combined knowledge mining model. By utilizing this ML model, banking authorities could better match their services to the needs of their clients, lower risk factors and increase client happiness. Customers may also have access to the greatest facilities and more banking services than in the past. This study did not address the problem of balancing classes in the data; hence, the model employed balanced benchmark datasets, which could not correlate to reality.

Sitharamulu et al. (2024) Deer Hunting Optimization (DHO) and Sunflower Optimization (SFO) Better precision value Clustering large data has become more complex due to their size

Navin and Krishnan (2024) Fuzzy Logic in DSMSVD Improving robustness and generalization More complexity

Author [Citation]	Methodologies	Features	Challenges
Mirashk et al. (2019)	RNNs to create a model for forecasting POS user activity	The created model was validated using actual data from one of Iran's private banks, and the projected outcomes were around 87% accurate	Need to recognize the difference between the predicted and natural behavior of customers
Raj Kannan et al. (2021)	CBA-MMP	To efficiently assess the patterns of consumer behavior, the suggested MMP-CBA combines the strengths of the DT algorithm with MLNN.	However, its analytical capabilities appeared to be weaker because it used ML technology to operate.
Argyris et al. (2020)	Visual congruence: DL algorithms for automatic image classification.	The study not only used DL algorithms for image classification, but it also suggested an ensemble method as a reliable strategy to support future academics in their efforts to delve into the world of visual content that is shared on social media.	It exclusively examined Instagram to show how aesthetic consistency increases consumer interaction with brands.
Tao and Zhou (2023)	DL-based time–series analysis.	The CNN-LSTM model had an accuracy of above 81%, making it useful for practical business closure prediction.	Since its conclusions were drawn on the analysis of data from a specific industry, the model was inapplicable to other industries.
Chaudhary et al. (2021)	ML-based social media consumer behavior model.	The installation of social media marketing and performance analysis was combined, and the data analyses were performed using a subpar instrument.	To operate on a real-time consumer basis, it was not implemented. This model's performance on real-time data would be very subpar.
Jeong (2021)	ML-based consumer preference identification model.	The proposed approach could be used by industry to create client-centered marketing strategies because it is adaptable, scalable and comprehensible for identifying significant drivers of consumer evaluations for various products in a particular industry.	If it could distinguish between real and fake reviews posted on various internet sites, it might provide a more accurate analysis of the data.
Yuan (2022)	Merging customer consumption behavior characteristics- algorithm	The study mines a huge quantity of consumer consumption data for a variety of forms of useful information, including customer preferences and consumption patterns	Have to utilize and larger number of samples to prove the efficiency of the algorithms
Li et al. (2022)	Tourism Demand Analysis Framework	The topic models and sentiment analysis methods suggested in this paper were used to extract sentiments, and the result was a social listening tool for predicting and managing tourism demand.	Since the study was restricted to a specific location and relied on data from just one social media platform, Twitter, its conclusions regarding the elements that influence demand needed to be more broadly generalized.
Meena and Kumar (2022)	LDA-based framework – text analysis, topic extraction, and sentiment analysis.	It normally looked into subjects including social responsibility, delivery procedures, perceived consumer responsibility and COVID-19 financial impacts and it offered better-studied results.	It was impossible to ascertain how people's sentiments related to the performance of OFD enterprises on the stock market throughout the COVID-19 period.
Abedin et al. (2023)	Combined knowledge mining model.	By utilizing this ML model, banking authorities could better match their services to the needs of their clients, lower risk factors and increase client happiness. Customers may also have access to the greatest facilities and more banking services than in the past.	This study did not address the problem of balancing classes in the data; hence, the model employed balanced benchmark datasets, which could not correlate to reality.
Sitharamulu et al. (2024)	Deer Hunting Optimization (DHO) and Sunflower Optimization (SFO)	Better precision value	Clustering large data has become more complex due to their size
Navin and Krishnan (2024)	Fuzzy Logic in DSMSVD	Improving robustness and generalization	More complexity

2.1 Research Gap

A review of existing works on consumer behavior analytics reveals a growing reliance on machine learning (ML) and deep learning (DL) approaches, such as RNNs, CNN-LSTM, fuzzy logic, optimization algorithms, and hybrid models. These methods have been applied to various domains such as banking, retail, tourism, and social media analytics. However, most suffer from common drawbacks such as limited scalability, poor generalizability across industries, difficulty in handling imbalanced datasets, and lack of interpretability. Some methods effectively use ensemble learning or sentiment analysis, while others rely on single-platform datasets, which restrict broader applicability. Few existing works integrate distributed frameworks like MapReduce with explainable models such as fuzzy rule-based systems. These gaps underscore the need for a scalable, interpretable, and generalizable framework like the proposed IFCBMC model.

3. Proposed Framework for Consumer Behavior Analytics via Fuzzy Rule-Based Classification

Consumer choices are influenced by a variety of circumstances, which lead to intended, impulsive in nature and unforeseen purchases. The behavior of the consumer is notable by the organizations or other marketing platforms to analyze their participation and facilitation towards the goods. This study uses a fuzzy rule-based classification model with data mapping and bagging techniques in the MapReduce scheme to deliver consumer behavior insights. This architecture illustrating the preprocessing, entropy-based feature extraction, and fuzzy ensemble classification phases within a distributed MapReduce environment. Preprocessing, feature extraction, and the prediction procedure are the three phases of the implementation.

Initially, the input big data D involve the behavior of the customers, like Income, Kid home, Teen home, etc., which are preprocessed by a data cleaning process and a modified normalization process.

The MapReduce framework is used to handle big data, where the feature extraction process is carried out from the normalized preprocessed data $D_{P}$ under the training dataset. Here, the features, namely, Improved entropy $F_{E}$ and Correlation $F_{C}$ are extracted in the mapper phase. The reducer combines all the features from the mappers.

The prediction model is an enhanced fuzzy rule-based classification model that uses a mapping function from the testing dataset and a bagging technique from the training dataset to enhance efficiency. Figure 1 displays the overall prediction model.

Figure 1.

Overview of IFCBMC Framework.

3.1 Preprocess the Data via Data Cleaning and Modified Normalization

Data preprocessing is a significant phase in the process of data mining. It defines the data-preparing processes for examination by cleansing, adapting and integrating it. The purpose of data preprocessing is to improve the data's quality and appropriateness for the specific data mining process. According to this work, the data cleaning and modified normalization process is carried out under the preprocessing phase, which is explained as follows.

3.1.1 Data Cleaning

Initially, the input data D are subjected to data cleaning, which is also known as data cleansing. This includes classifying and modifying errors or variations in the data, such as lost values, outliers, and duplicates. Therefore, data cleaning is to find the easiest way to rectify the quality issues and also convert the data into standardized numeric data which is easier to manipulate. The cleaned input data is then placed through an enhanced normalization step once the data-cleaning phase is complete.

3.1.2 Modified Normalization

Initially, the input data undergo cleaning to remove noise, inconsistencies, and missing values. Following this, a modified normalization process is applied to standardize the data for machine learning tasks. This normalization approach combines min-max normalization with the Tanh estimator to enhance data scaling and stability.

(i) Min-max normalization

The min-max normalization (Horng et al., 2009) is termed a linear transformation, which preserves the relationships between the original data values. The data are usually in the range of (0,1) and their calculation is expressed in Equation (1).

\begin{aligned} M i n max (x^{'}) = \frac{x - min (x)}{max (x) - min (x)} \end{aligned}

(1)

Where $M i n max (x^{'})$ is the normalized value, x is the original value, $max (x)$ is the maximum raw value, and $min (x)$ is the minimum raw value.

(ii) Tanh Normalization

The outcome from the min-max normalization is applied to the Tanh estimators (Prasetyo et al., 2020), regarded as being a more effective and reliable normalization method. It produces values ranging from −1 to 1. The normalized value $x^{″}$ (the preprocessed result $D_{P}$ ) of the value x of $M i n max (x^{'})$ an attribute of the input data set can be calculated by the following Equation (2), where, $k = 0.5$ , the mean represents the contra harmonic mean. Thereby, the final preprocessed normalized data $D_{P}$ are subjected to handling with the MapReduce framework.

\begin{aligned} D_{P} = x^{″} = \frac{1}{2} {\tanh [k (\frac{x - m e a n (M i n max (x^{'}))}{s t d (M i n max (x^{'}))})] + 1} \end{aligned}

(2)

3.2 MapReduce Framework for Bigdata Handling: A Feature Extraction Process

This study handles big data, or customer behavior-related data, using the MapReduce architecture. According to the MapReduce framework (del Rio et al., 2015), the Mapper phase performs feature extraction, which extracts the Improved entropy and Correlation features. Similarly, the features gathered from the several mappers are combined in the Reducer step. The MapReduce framework's extracted features are displayed in Figure 2.

Figure 2.

MapReduce Framework for Feature Extraction. This Figure Depicts How Consumer Data are Split Across Mapper Nodes for Entropy and Correlation Feature Extraction and Then Aggregated by Reducers. This Process Enables Scalable and Parallel Processing for High-Dimensional Big Data.

The map and the reduce functions are the two main components of the MapReduce concept. In general, the mapping function analyzes the input data in the initial phase, producing some intermediate outcomes. Accordingly, this work assists the mapper function for feature extraction. The feature extraction process is carried out for the normalized preprocessed data, $D_{P}$ . Those intermediate outcomes are then passed to the reducing function in the second phase, which effectively merges the intermediate outcomes to get the result. The final feature set is thus determined by the reduction by combining all of the features that were taken from the mappers.

3.2.1 Map Function

The input data are split into a variable n number of mappers in the map function. The map function takes training normalized data as input and outputs a group of intermediate data. In this paper, improved entropy and correlation features are extracted from the map function. This enhancement allows us to capture nuances in feature distributions more effectively, thereby improving the discriminatory power of extracted features. Features exhibiting high correlation coefficients are prioritized for inclusion in subsequent analysis, aiming to enhance model interpretability and predictive accuracy.

3.2.1.1 Improved Entropy $F_{E}$

In information theory, entropy is more precisely the anticipated quantity of shared information (Trovati et al., 2020) across all its possible combinations. In this paper, an improved entropy-based feature is proposed, which is the combination of modified Deng entropy and Belief entropy (Zhao et al., 2019).

(a) Modified Deng entropy

Deng entropy can estimate an unidentified degree more effectively than other uncertain measurements. However, Modified Deng entropy is suggested in this work, which offers efficiency because it more thoroughly considers all potential possibilities. In Equation (3), M denotes the mass function, $| S |$ represents the cardinality of the body of evidence (BOE), $| A |$ and represents the cardinality of the focal element A. The exponential factor $e^{\frac{\log_{2} (2^{| A |} - 1)}{| S |}}$ in the modified Deng entropy denotes the uncertain information in a BOE, which is significantly reduced as compared to the conventional Deng entropy.

\begin{aligned} \begin{array}{l} M H_{d} (M) = [\frac{{[\sum_{A \subseteq S} M (A) \log_{2} (2^{| A |} - 1) + e^{\frac{\log_{2} (2^{| A |} - 1)}{| S |}}] - [\sum_{A \subseteq S} M (A) \log_{2} M (A) + e^{\frac{\log_{2} (2^{| A |} - 1)}{| S |}}]}}{e^{\log_{2} 2^{| A |} M (A)}}] \end{array} \end{aligned}

(3)

(b) Belief Entropy

Additional information in BOE is addressed by the belief entropy, such as the scale of FOD, indicated as $| S |$ and the relative importance of a focused element about FOD, designated as $\frac{| A | - 1}{| S |}$ . Where A is the focal element (subset of frame discernment, FOD), s denotes the cardinality (number of elements in FOD), the cardinality of focal element A, and m(A) indicates the basic probability assignment for focal element A. According to Equation (4), the belief entropy is determined. Adding Equations (3) and (4) will improve the entropy measure. The enhanced entropy, which is the result of combining the modified Deng entropy and belief entropy, is shown in Equations (5) and (6).

\begin{aligned} H_{d} (M) & = - (\sum_{A \subseteq S} M (A) \log_{2} (\frac{M (A)}{2^{| A |} - 1} . e^{\frac{| A | - 1}{| S |}})) \end{aligned}

(4)

\begin{aligned} F_{E} & = [M H_{d} (M) + H_{d} (M)] \end{aligned}

(5)

\begin{aligned} F_{E} & = I H_{d} (M) = {\begin{cases} \frac{[\begin{array}{l} (\sum_{A \subseteq S} M (A) \log_{2} (2^{| A |} - 1) + e^{\frac{\log_{2} (2^{| A |} - 1)}{| S |}}) - \\ (\sum_{A \subseteq S} M (A) \log_{2} M (A) + e^{\frac{\log_{2} (2^{| A |} - 1)}{| S |}}) \end{array}]}{[e^{\log_{2} 2^{| A |} M (A)}]} + \\ (- \sum_{A \subseteq S} M (A) \log_{2} (\frac{M (A)}{2^{| A |} - 1} . e^{\frac{| A | - 1}{| S |}})) \end{cases}} \end{aligned}

(6)

3.2.1.2 Correlation

F_{C}

The level of correlation between the two variables can be identified through a correlation analysis by the Pearson correlation coefficient. It is impossible to directly compute the Pearson correlation coefficient among the two-component values of the combined body. First, it is necessary to gather the samples that represent the two index factors, determine the Pearson correlation coefficient, and then apply the Pearson correlation coefficient among the samples of the two index factors to calculate the Pearson correlation coefficient among the two index variables. Using the obtained specimens, the sample correlation coefficient is utilized to calculate the entire correlation coefficient to derive the Pearson correlation coefficient among the total bodies of the two index factors.

There is no interaction between the two variables if the value is 0. As shown in Equation (6), the Pearson correlation coefficient is determined. Where, $C_{c}$ is the Pearson correlation coefficient, $x_{j}$ and $y_{j}$ denote the values that were found of the dataset class, $\bar{x}$ indicates the Harmonic mean values of the features under consideration, $\bar{y}$ stands for the arithmetic mean values, and $C_{c}$ indicates −1 for a negative correlation, + 1 for a positive correlation, and 0 for no correlation. This correlation study $F_{C}$ only considers a linear link.

\begin{aligned} F_{C} = C_{c} = \frac{\sum_{j = 1}^{n} (x_{j} - \bar{x}) (y_{j} - \bar{y})}{\sqrt{\sum_{j = 1}^{n} {(x_{j} - \bar{x})}^{2}} \sqrt{\sum_{j = 1}^{n} {(y_{j} - \bar{y})}^{2}}} \end{aligned}

(7)

3.2.2 Reduce Function

The features from the map function are shuffled and reduced in the reduce function, and then ultimately combined to create the final extracted features. The parameters of the classifier details are tabulated in Table 2. Hence, the total feature extraction set is represented as $F = [F_{E} F_{C}]$ .

Table 2.
Parameter of Classifiers.

Classifiers Parameters

LSTM Loss – sparse categorical cross-entropy

Activation: softmax

Epochs – 50

Batch size = 100

CNN Loss – sparse categorical cross-entropy

Epochs – 50

Batch size – 100

DCNN Loss – sparse categorical cross-entropy

Epochs – 50

Batch size – 100

Bi-GRU Loss – categorical cross-entropy

Epochs – 50

Batch size – 100

RNN Loss – sparse categorical cross-entropy

Epochs – 50

Batch size – 100

DNN Loss – sparse categorical cross-entropy

Epochs – 50

Batch size – 100

Classifiers	Parameters
LSTM	Loss – sparse categorical cross-entropy
Activation: softmax
Epochs – 50
Batch size = 100
CNN	Loss – sparse categorical cross-entropy
Epochs – 50
Batch size – 100
DCNN	Loss – sparse categorical cross-entropy
Epochs – 50
Batch size – 100
Bi-GRU	Loss – categorical cross-entropy
Epochs – 50
Batch size – 100
RNN	Loss – sparse categorical cross-entropy
Epochs – 50
Batch size – 100
DNN	Loss – sparse categorical cross-entropy
Epochs – 50
Batch size – 100

3.3 Prediction Process: Improved Fuzzy Rule-Based Classification Model

In the Prediction phase, the extracted features F are subjected to the Improved fuzzy rule-based classification model (Trawinski et al., 2011) that adopts the Improved bagging and mapping criteria, which enhances the prediction process by considering the behavior of consumers. It specifies the fuzzy rule-based classification algorithm used in the IFCBMC. To improve robustness and generalization, we employ a bagging approach where the dataset is randomly sampled with replacement into multiple bags. Each bag serves as a training set for a distinct fuzzy rule-based classifier, allowing us to capture diverse patterns in customer behavior. Bagged classifiers are integrated using a voting mechanism in the ensemble phase, enhancing predictive accuracy and reducing overfitting. Mappers extract features from distributed datasets and pass them to reducers, where features are combined and aggregated. Fuzzy logic is particularly well-suited for modeling human behavior, as it allows for approximate reasoning and handles ambiguity effectively—ideal for interpreting variables like purchase intent, income brackets, and household dynamics. The MapReduce framework is employed to ensure scalability and efficiency in processing massive datasets by parallelizing feature extraction and aggregation across distributed computing nodes. Within this, the inclusion of Improved Entropy (combining Modified Deng Entropy and Belief Entropy) and correlation-based features provides robust uncertainty modeling and relevance scoring for high-impact features. To further enhance generalization and reduce overfitting, the Bagging ensemble approach is applied, allowing multiple classifiers trained on different subsets to improve prediction stability and accuracy. The overall prediction procedure is explained as follows.

Initially, the preprocessed, normalized dataset is separated into two groups: the testing dataset and the training dataset. This is done by dividing the normalized dataset into two categories: 50% of the normalized dataset is used for training or 50% for testing.

The Mapper Reducer system is used to extract features from the training dataset. Improved entropy and correlation are extracted as part of the feature extraction process in this mapper step. The features from different mappers are combined in the Reducer step. to obtain the extracted features’ final result.

After that, an ensemble of fuzzy rule-based classifiers is created using bagging. Here, distinct training sets for each basic classifier are created by randomly selecting portions of data with replacements. The process of bagging involves averaging numerous classifiers, each of which is adjusted to random samples that correspond to the sample distribution of the training set, in order to minimize the classification variance. The final model is significantly altered by even slight modifications to the training set; bagging works best with these unstable classifiers. Additionally, given the limited number of samples in the dataset, it is recommended. Additionally, bagging enables learners in the ensemble to train independently and in simultaneously. Simultaneously, the 50% testing dataset from the normalized data is taken into the mapping function. Here, the testing samples are split into multiple related mappers in the map function. Then each mapper's data is taken as input into the final feature set of classifiers.

In the fuzzy rule-based classification model, both the mapping and bagging approaches from the testing and training datasets are combined to get the final set of classifiers. Here, 72 built-in rules are involved in the fuzzy classifiers. Then the classifier data are shuffled to get the final output in the Reducer phase.

The original classification set of scores will be obtained. For each score improved weight factor $W_{j}$ is added, which is calculated as per Equation (8). Where, $W_{j}$ denotes the weight for each class ( $j$ signifies the class), $n_S a m p l e s$ denotes the total number of samples or rows in the dataset, $n_c l a s s e s$ represents the total number of unique classes in the target, and $n_S a m p l e s_{j}$ denotes the total number of rows of the respective class. Next Mapping procedure is performed, and finally the final prediction.

\begin{aligned} W_{j} = (n_S a m p l e s / (n_c l a s s e s * n_S a m p l e s_{j}) * \frac{1}{1 + e^{- (n_c l a s s e s * n_S a m p l e s_{j})}}) \end{aligned}

(8)

Thus, the prediction outcome of the customer behavior is generated based on seven label categories. Examples of these are extremely low, low, medium-low, medium, medium-high, high, and very high. Figure 3 shows the overall prediction process of the IFCBMC model.

Figure 3.

Prediction Structure of the Improved Fuzzy Classification With Bagging and MapReduce Coordination (IFCBMC) Model. It Shows How Extracted Features are Used to Train Multiple Fuzzy Rule-Based Classifiers With Bagging and How Predictions are Aggregated, Enhancing Accuracy and Robustness.

4 Results Analysis on IFCBMC for Consumer Behavior Analytics Framework

4.1 Experimental Setup

The proposed consumer behavior analytics was simulated using PYTHON. The Python version utilized for this implementation was “Python 3.7” and the processor utilized was “Intel(R) Core(TM) i5-4210U CPU @ 1.70 GHz 1.70 GHz.” Moreover, the analysis of consumer behavior classification was conducted using the Consumer Buying Behavior Analysis dataset (Pratap, 2022).

4.2 Dataset Description

This dataset incorporates three distinct categories of attributes: People, Products, Promotion, and Place. Within the People category, you'll find attributes like ID, Year of Birth, Education, Income, and more. The Products category encompasses attributes such as MntWines, MntFruits, MntGoldProds, and so forth. The Promotion category includes features like NumDealsPurchases, AcceptedCmp1, AcceptedCmp2, Response, and others. Finally, the Place category comprises attributes like NumWebPurchases, NumCatalogPurchases, NumStorePurchases, and so on.

4.3 Performance Analysis

Both the standard approaches and the IFCBMC approach underwent the categorization examination. Furthermore, the IFCBMC method was compared to state-of-the-art approaches such as RNN [24] and DNN [28], as well as traditional methods including LINKNET, SQUEEZENET, LSTM, SVM, CNN, DCNN, and Bi-GRU.

4.4 Comparative Study on Positive Metrics

Figure 4 illustrates the comparative evaluation of positive metrics between the IFCBMC method and LINKNET, SQUEEZENET, LSTM, SVM, CNN, DCNN, Bi-GRU, RNN [28], and DNN [28] in the context of consumer behavior analytics. To ensure the precise classification of consumer behavior, the model must achieve higher accuracy scores. Particularly, for the training rate 70, the IFCBMC scored an accuracy rate of 92.492, even though the traditional schemes attained minimal accuracy values, notably, LINKNET = 79.1, SQUEEZENET = 81.12, LSTM = 80.174, SVM = 81.236, CNN = 82.465, DCNN = 85.357, Bi-GRU = 83.591, RNN [24] = 84.753, and DNN [28] = 86.846, correspondingly. In particular, the precision acquired by the IFCBMC approach is much higher than LINKNET, SQUEEZENET, LSTM, SVM, CNN, DCNN, Bi-GRU, RNN [24], and DNN [28] in the entire training rates.

Figure 4.

Performance Comparison Based on Positive Evaluation Metrics (Accuracy, Precision, Sensitivity). This Bar Graph Compares the IFCBMC Model to Several Existing Classifiers, Demonstrating its Superior Performance Across Key Metrics Critical for Successful Customer Behavior Classification.

The analysis of sensitivity and specificity evaluations for IFCBMC and existing models is presented in Figure 4c and d. Further, the sensitivity of the IFCBMC methodology is 88.932 (training rate = 90), even though the LINKNET is 62.63, SQUEEZENET is 65.124, LSTM is 62.541, SVM is 61.836, CNN is 60.514, DCNN is 64.947, Bi-GRU is 58.792, RNN [24] is 59.488, and DNN [28] is 63.814, correspondingly. In summary, the remarkable performance of the IFCBMC approach provides strong assurance of its capability to effectively classify consumer behavior. This capability is achieved by integrating modified normalization, enhanced entropy, improved fuzzy rule-based classification, and final classification accomplished via improved bagging and mapping.

4.5 Comparative Study on Negative Metric

The assessment on IFCBMC is contradicted by LINKNET, SQUEEZENET, LSTM, SVM, CNN, DCNN, Bi-GRU, RNN [24], and DNN [28] with regard to the negative metric for consumer behavior analysis, as shown in Figure 5. At a training rate of 90, the FNR achieved by the IFCBMC approach is notably lower at 14.382 compared to LINKNET, SQUEEZENET, LSTM, SVM, CNN, DCNN, Bi-GRU, RNN [24], and DNN [28]. Moreover, the FPR achieved with the IFCBMC method is significantly lower across all training rates. The least FPR rate attained using the IFCBMC is 3.218 at the training rate of 80, whereas the LINKNET (7.653), SQUEEZENET (7.854), LSTM (8.137), SVM (7.842), CNN (8.288), DCNN (8.198), Bi-GRU (8.934), RNN [24] (7.761), and DNN [28] (8.478) recorded greater FPR values. The implementation of the IFCBMC technique has been credited with the evaluation's remarkable achievement in consumer behavior analysis. This improvement can be ascribed to the use of enhanced bagging and mapping algorithms, changed normalization, enhanced entropy, and enhanced fuzzy rule-based categorization.

Figure 5.

Performance Comparison Based on Negative Evaluation Metrics (False Positive Rate and False Negative Rate). Illustrates How the Proposed Model Significantly Reduces Error Rates Compared to Conventional Methods, Indicating More Reliable Classification.

4.6 Comparative Study on Other Metrics

The other metric analysis on IFCBMC is contradicted by LINKNET, SQUEEZENET, LSTM, SVM, CNN, DCNN, Bi-GRU, RNN [24], and DNN [28] for consumer behavior analysis is exposed in Figure 6. Additionally, the NPV of the IFCBMC approach is 92.485 (training rate = 90), whilst the LINKNET is 81.325, SQUEEZENET is 82.685, LSTM is 86.139, SVM is 85.215, CNN is 84.382, DCNN is 87.956, Bi-GRU is 83.763, RNN [24] is 81.437, and DNN [28] is 82.519, correspondingly. Moreover, the highest MCC attained through the IFCBMC approach is 87.902, surpassing LINKNET, SQUEEZENET, LSTM, SVM, CNN, DCNN, Bi-GRU, RNN [24], and DNN [28]. Therefore, by demonstrating better values in other measures, the IFCBMC methodology outperformed previously used methodologies, demonstrating its incredible potential for correctly categorizing customer behavior.

Figure 6.

Evaluation of Other Classification Metrics (F1-Score, NPV, Specificity). This Figure Presents the Model's Balanced Performance Across Additional Key Metrics, Reinforcing Its General Effectiveness in Consumer Behavior Analysis.

4.7 Ablation Evaluation on IFCBMC

The ablation study on IFCBMC, a model with conventional entropy, a model with standard normalization, and a model without feature extraction for consumer behavior analysis are explained in Table 3. This process helps us better understand the distinct contributions these components make to the effectiveness of the IFCBMC system. Additionally, the traditional normalization model has a specificity of 0.870, the conventional entropy model has a specificity of 0.917, the IFCBMC technique has a specificity of 0.960, and the model without feature extraction has a specificity of 0.836. Furthermore, the FPR of the IFCBMC technique is 0.040, the traditional Normalization model is 0.130, the model without feature extraction is 0.164, and the conventional Entropy model is 0.083.

Table 3.
Ablation Assessment on IFCBMC, Model Without Feature Extraction, Model With Conventional Normalization, and Model With Conventional Entropy.

Model Without Model With Conventional Model With Conventional

Metrics Feature Extraction Normalization Entropy IFCBMC

Accuracy 0.782 0.812 0.857 0.931

NPV 0.835 0.869 0.917 0.960

FPR 0.164 0.130 0.083 0.040

Sensitivity 0.681 0.703 0.752 0.758

F-measure 0.463 0.468 0.501 0.758

FNR 0.319 0.297 0.248 0.242

Specificity 0.836 0.870 0.917 0.960

MCC 0.388 0.387 0.418 0.718

Precision 0.463 0.468 0.501 0.759

	Model Without	Model With Conventional	Model With Conventional
Accuracy	0.782	0.812	0.857	0.931
NPV	0.835	0.869	0.917	0.960
FPR	0.164	0.130	0.083	0.040
Sensitivity	0.681	0.703	0.752	0.758
F-measure	0.463	0.468	0.501	0.758
FNR	0.319	0.297	0.248	0.242
Specificity	0.836	0.870	0.917	0.960
MCC	0.388	0.387	0.418	0.718
Precision	0.463	0.468	0.501	0.759

4.8 Statistical Analysis of Accuracy

Table 4 describes the statistical assessment of IFCBMC is contradicted by LINKNET, SQUEEZENET, LSTM, SVM, CNN, DCNN, Bi-GRU, RNN [24], and DNN [28] for consumer behavior analysis. To ensure that highly precise calculations are produced, each approach is evaluated in detail. In terms of the greatest statistical metric, the IFCBMC technique achieves the highest accuracy with an astounding accuracy of 0.960. In contrast, the traditional approaches showed lower accuracy rates, specifically, LSTM with 0.862, LINKNET with 0.866, SQUEEZENET with 0.858, SVM with 0.863, DCNN with 0.860, Bi-GRU with 0.854, RNN [24] with 0.861, and DNN [28] with 0.860, respectively. Furthermore, the IFCBMC method achieved an accuracy rate of 0.938 when evaluated under the median statistical metric, while LSTM, SVM, CNN, DCNN, Bi-GRU, RNN [24], and DNN [28] all produced lower accuracy scores. This superior performance is attributed to the integration of fuzzy rule-based classification, entropy-based feature extraction, and bagging ensemble techniques within a MapReduce framework which together effectively handle data uncertainty, distribution, and high dimensionality.

Table 4.
Statistical Evaluation of Accuracy, Precision, Recall, and F-Measure.

Statistical Metrics Min Max Mean Median Standard Deviation

LSTM 0.863308 0.867985 0.865102 0.864557 0.001813

CNN 0.853954 0.863042 0.858379 0.858259 0.003264

LINKNET 0.862404 0.866709 0.864756 0.864955 0.001572

SQUEEZENET 0.85693 0.861607 0.858883 0.858498 0.001849

DCNN 0.856665 0.859906 0.858432 0.858578 0.001204

Bi-GRU 0.84375 0.853529 0.848467 0.848294 0.003538

RNN [24] 0.846301 0.860651 0.855641 0.857807 0.00565

DNN [28] 0.8446 0.860332 0.851974 0.851483 0.005586

SVM 0.853954 0.862564 0.857262 0.856266 0.003207

IFCBMC 0.918527 0.959821 0.938683 0.938191 0.015489

Precision

Methods Min Max Mean Median Standard deviation

LINKNET 0.522321 0.540179 0.529018 0.526786 0.00688

SQUEEZENET 0.518973 0.534226 0.527809 0.529018 0.005748

DCNN 0.498884 0.510417 0.505673 0.506696 0.00449

CNN 0.491071 0.521205 0.505487 0.504836 0.010832

SVM 0.491071 0.520089 0.501581 0.497582 0.011031

BiGRU 0.455357 0.488095 0.470796 0.469866 0.011851

RNN 0.464286 0.513393 0.495908 0.502976 0.019212

LSTM 0.5 0.517857 0.507254 0.50558 0.006947

DNN 0.456845 0.513393 0.483073 0.481027 0.020113

IFCBMC 0.918527 0.959821 0.938683 0.938191 0.015489

Recall

LINKNET 0.681545 0.697778 0.687784 0.685907 0.006296

SQUEEZENET 0.678395 0.693432 0.686582 0.687251 0.005482

DCNN 0.647826 0.662556 0.655848 0.656505 0.005467

BiGRU 0.589333 0.633581 0.610632 0.609808 0.016012

CNN 0.635556 0.676812 0.655627 0.65507 0.01482

LSTM 0.649034 0.670222 0.657893 0.656159 0.00838

RNN 0.600889 0.665924 0.643213 0.65302 0.025606

SVM 0.635556 0.67461 0.650549 0.646015 0.014554

DNN 0.593016 0.664444 0.62651 0.624289 0.025363

IFCBMC 0.714604 0.857778 0.784662 0.783132 0.053717

F-MEASURE

LINKNET 0.521933 0.538976 0.5284 0.526346 0.006584

SQUEEZENET 0.518684 0.533829 0.527195 0.528133 0.005606

DNN 0.456506 0.512249 0.482501 0.480624 0.01981

SVM 0.489978 0.519509 0.501001 0.497258 0.011109

CNN 0.489978 0.520915 0.504907 0.504368 0.011116

BiGRU 0.454343 0.487732 0.470256 0.469474 0.012083

RNN 0.463252 0.512821 0.495343 0.50265 0.019455

LSTM 0.499628 0.516704 0.506662 0.505158 0.006692

DCNN 0.498606 0.510037 0.505085 0.505848 0.004337

IFCBMC 0.715003 0.859688 0.785605 0.783864 0.054267

Statistical Metrics	Min	Max	Mean	Median	Standard Deviation
LSTM	0.863308	0.867985	0.865102	0.864557	0.001813
CNN	0.853954	0.863042	0.858379	0.858259	0.003264
LINKNET	0.862404	0.866709	0.864756	0.864955	0.001572
SQUEEZENET	0.85693	0.861607	0.858883	0.858498	0.001849
DCNN	0.856665	0.859906	0.858432	0.858578	0.001204
Bi-GRU	0.84375	0.853529	0.848467	0.848294	0.003538
RNN [24]	0.846301	0.860651	0.855641	0.857807	0.00565
DNN [28]	0.8446	0.860332	0.851974	0.851483	0.005586
SVM	0.853954	0.862564	0.857262	0.856266	0.003207
IFCBMC	0.918527	0.959821	0.938683	0.938191	0.015489
Precision
Methods	Min	Max	Mean	Median	Standard deviation
LINKNET	0.522321	0.540179	0.529018	0.526786	0.00688
SQUEEZENET	0.518973	0.534226	0.527809	0.529018	0.005748
DCNN	0.498884	0.510417	0.505673	0.506696	0.00449
CNN	0.491071	0.521205	0.505487	0.504836	0.010832
SVM	0.491071	0.520089	0.501581	0.497582	0.011031
BiGRU	0.455357	0.488095	0.470796	0.469866	0.011851
RNN	0.464286	0.513393	0.495908	0.502976	0.019212
LSTM	0.5	0.517857	0.507254	0.50558	0.006947
DNN	0.456845	0.513393	0.483073	0.481027	0.020113
IFCBMC	0.918527	0.959821	0.938683	0.938191	0.015489
Recall
LINKNET	0.681545	0.697778	0.687784	0.685907	0.006296
SQUEEZENET	0.678395	0.693432	0.686582	0.687251	0.005482
DCNN	0.647826	0.662556	0.655848	0.656505	0.005467
BiGRU	0.589333	0.633581	0.610632	0.609808	0.016012
CNN	0.635556	0.676812	0.655627	0.65507	0.01482
LSTM	0.649034	0.670222	0.657893	0.656159	0.00838
RNN	0.600889	0.665924	0.643213	0.65302	0.025606
SVM	0.635556	0.67461	0.650549	0.646015	0.014554
DNN	0.593016	0.664444	0.62651	0.624289	0.025363
IFCBMC	0.714604	0.857778	0.784662	0.783132	0.053717
F-MEASURE
LINKNET	0.521933	0.538976	0.5284	0.526346	0.006584
SQUEEZENET	0.518684	0.533829	0.527195	0.528133	0.005606
DNN	0.456506	0.512249	0.482501	0.480624	0.01981
SVM	0.489978	0.519509	0.501001	0.497258	0.011109
CNN	0.489978	0.520915	0.504907	0.504368	0.011116
BiGRU	0.454343	0.487732	0.470256	0.469474	0.012083
RNN	0.463252	0.512821	0.495343	0.50265	0.019455
LSTM	0.499628	0.516704	0.506662	0.505158	0.006692
DCNN	0.498606	0.510037	0.505085	0.505848	0.004337
IFCBMC	0.715003	0.859688	0.785605	0.783864	0.054267

4.9 K-Fold Analysis

The K-fold analysis is compared with conventional approaches like LINKNET, SQUEEZENET, LSTM, SVM, CNN, DCNN, Bi-GRU, RNN [24], and DNN [28] for consumer behavior analytics as summarized in Table 5. A common approach to evaluate the robustness and efficacy of prediction models in machine learning and statistics is referred as K-fold cross-validation. Mainly, the IFCBMC scheme achieved an NPV of 0.967, significantly surpassing the much lower NPV values obtained by LSTM (0.891), SVM (0.892), CNN (0.893), Bi-GRU (0.887), RNN [24] (0.379), and DNN [28] (0.389). The high NPV and consistent K-fold performance confirm that the IFCBMC model is not only accurate but also highly dependable for real-world consumer behavior analysis.

Table 5.
Assessment on K-fold.

Methods Accuracy Sensitivity Specificity Precision F_measure MCC NPV FPR FNR

LINKNET 0.86588 0.680512 0.92192 0.681696 0.681104 0.602853 0.921577 0.07808 0.469488

SQUEEZENET 0.864477 0.675612 0.921102 0.676786 0.676198 0.597129 0.920759 0.078898 0.474388

LSTM 0.813967 0.620423 0.891626 0.620813 0.620618 0.566348 0.891295 0.108374 0.379577

SVM 0.814605 0.621723 0.891999 0.622115 0.621919 0.567835 0.891667 0.108001 0.378277

CNN 0.817156 0.626922 0.893487 0.627324 0.627123 0.573783 0.893155 0.106513 0.373078

DCNN 0.81486 0.622243 0.892147 0.622636 0.622439 0.56843 0.891815 0.107853 0.377757

BiGRU 0.806569 0.605347 0.887309 0.605708 0.605528 0.5491 0.886979 0.112691 0.394653

RNN 0.814349 0.621203 0.89185 0.621594 0.621399 0.567241 0.891518 0.10815 0.378797

DNN 0.809503 0.611326 0.889021 0.611698 0.611512 0.55594 0.88869 0.110979 0.388674

Proposed 0.944196 0.804009 0.967622 0.805804 0.804905 0.772348 0.967262 0.032378 0.195991

Methods	Accuracy	Sensitivity	Specificity	Precision	F_measure	MCC	NPV	FPR	FNR
LINKNET	0.86588	0.680512	0.92192	0.681696	0.681104	0.602853	0.921577	0.07808	0.469488
SQUEEZENET	0.864477	0.675612	0.921102	0.676786	0.676198	0.597129	0.920759	0.078898	0.474388
LSTM	0.813967	0.620423	0.891626	0.620813	0.620618	0.566348	0.891295	0.108374	0.379577
SVM	0.814605	0.621723	0.891999	0.622115	0.621919	0.567835	0.891667	0.108001	0.378277
CNN	0.817156	0.626922	0.893487	0.627324	0.627123	0.573783	0.893155	0.106513	0.373078
DCNN	0.81486	0.622243	0.892147	0.622636	0.622439	0.56843	0.891815	0.107853	0.377757
BiGRU	0.806569	0.605347	0.887309	0.605708	0.605528	0.5491	0.886979	0.112691	0.394653
RNN	0.814349	0.621203	0.89185	0.621594	0.621399	0.567241	0.891518	0.10815	0.378797
DNN	0.809503	0.611326	0.889021	0.611698	0.611512	0.55594	0.88869	0.110979	0.388674
Proposed	0.944196	0.804009	0.967622	0.805804	0.804905	0.772348	0.967262	0.032378	0.195991

4.10 Sensitivity Analysis

Table 6 represents the sensitivity analysis of various metrics. The percentage of accurately identified occurrences among all events is known as accuracy. In this case, Conventional Normalization (0.812064) yields the highest accuracy, indicating it effectively improves the overall predictive performance compared to other normalization methods. Different normalization techniques affect how the data are prepared for analysis, influencing the performance of classification models. Conventional Normalization consistently shows strong performance across multiple metrics, suggesting it effectively prepares the data for accurate classification. Conventional Normalization, which often refers to standard min-max scaling or similar methods, shows robust performance across various metrics, making it a reliable choice for many applications in consumer behavior analytics.

Table 6.
Sensitivity Analysis of Various Metrics.

Proposed Proposed With Proposed With Proposed Proposed With

Without z-score_ Robust_ With Conventional_

Metrics Normalization Normalization Normalization l1_Normalization Normalization Proposed

MCC 0.390037 0.395609 0.391894 0.389503 0.387111 0.718057

Specificity 0.840589 0.85585 0.845676 0.857656 0.869635 0.959811

FPR 0.159411 0.14415 0.154324 0.142344 0.130365 0.040189

F_measure 0.465177 0.472364 0.467572 0.467578 0.467584 0.758364

Sensitivity 0.685884 0.699079 0.690282 0.696627 0.702972 0.757801

Precision 0.465523 0.472716 0.46792 0.467926 0.467932 0.758929

NPV 0.840381 0.855638 0.845467 0.857443 0.86942 0.959573

Accuracy 0.78683 0.800935 0.791531 0.801798 0.812064 0.93091

FNR 0.314116 0.300921 0.309718 0.303373 0.297028 0.242199

	Proposed	Proposed With	Proposed With	Proposed	Proposed With
MCC	0.390037	0.395609	0.391894	0.389503	0.387111	0.718057
Specificity	0.840589	0.85585	0.845676	0.857656	0.869635	0.959811
FPR	0.159411	0.14415	0.154324	0.142344	0.130365	0.040189
F_measure	0.465177	0.472364	0.467572	0.467578	0.467584	0.758364
Sensitivity	0.685884	0.699079	0.690282	0.696627	0.702972	0.757801
Precision	0.465523	0.472716	0.46792	0.467926	0.467932	0.758929
NPV	0.840381	0.855638	0.845467	0.857443	0.86942	0.959573
Accuracy	0.78683	0.800935	0.791531	0.801798	0.812064	0.93091
FNR	0.314116	0.300921	0.309718	0.303373	0.297028	0.242199

4.11 ROC Curve Analysis

Figure 7 exposes the ROC curve evaluation on IFCBMC compared with LSTM, SVM, CNN, DCNN, Bi-GRU, RNN [24], and DNN [28] for consumer behavior analysis. The ROC curve is produced by graphing the true positive rate against the false positive rate, particularly at the 70% training rate. Likewise, with the false positive rate set at 1.0, the IFCBMC approach achieved a true positive rate of 0.974, outperforming the lower true positive rates achieved by LSTM, SVM, CNN, DCNN, Bi-GRU, RNN [24], and DNN [28]. The trapezoidal rule, which numerically integrates the ROC curve points, was used to determine the Area under the Curve (AUC). Higher values denote stronger discrimination performance of the classifier over all feasible thresholds. The range of AUC values is 0 to 1. Better performance of the classifier in differentiating between classes across all thresholds is indicated by AUC values nearer 1. Thus, the positive findings of the ROC curve evaluation demonstrate the ability of the IFCBMC approach to correctly categorize consumer behavior through the integration of improved entropy, fuzzy rule-based classification, modified normalization, and, finally, improved bagging and mapping rules that develop the performance of classification.

Figure 7.

Receiver Operating Characteristic (ROC) Curve Comparison. The ROC Curves Demonstrate the True Positive Rate Versus False Positive Rate for the IFCBMC and Benchmark Models. The Area Under the Curve (AUC) Indicates the Proposed Model's Strong Discriminative Ability.

4.12. Proposed Model Using Different Entropy-Based Feature Extraction Techniques

Table 7 represents the analysis of the proposed consumer behavior analytics model using different entropy-based feature extraction methods, Shannon entropy, Belief entropy, Deng entropy, and Improved Deng entropy, highlighting the superior performance of the Improved Deng entropy approach across nearly all evaluation metrics. Its maximum accuracy (0.9247), specificity (0.9562), precision (0.7372), F-measure (0.7367), and Matthews Correlation Coefficient (MCC) of 0.6928 are significant; these results demonstrate its strong classification capacity, particularly in accurately detecting both positive and negative occurrences. While sensitivity is slightly lower than with belief entropy, the trade-off is justified by the overall boost in precision and model balance. These results demonstrate that incorporating Improved Deng entropy significantly enhances the discriminative power of features, resulting in more accurate and reliable consumer behavior predictions.

Table 7.
Proposed Model Using Different Entropy-Based Feature Extraction Techniques.

Proposed With Proposed With Proposed With Proposed With Only

Only Shannon Only Belief Only Deng Improved Deng

Metrics Entropy Entropy Entropy Entropy Features

Accuracy 0.852649 0.864421 0.871487 0.924718

Sensitivity 0.706144 0.741122 0.730386 0.736203

Specificity 0.89539 0.915809 0.914723 0.956185

Precision 0.597425 0.557459 0.61343 0.737165

F_measure 0.647251 0.636302 0.666819 0.736684

MCC 0.537311 0.48524 0.552584 0.692765

NPV 0.895187 0.915582 0.914497 0.955977

FPR 0.10461 0.084191 0.085277 0.043815

FNR 0.293856 0.258878 0.269614 0.263797

	Proposed With	Proposed With	Proposed With	Proposed With Only
Accuracy	0.852649	0.864421	0.871487	0.924718
Sensitivity	0.706144	0.741122	0.730386	0.736203
Specificity	0.89539	0.915809	0.914723	0.956185
Precision	0.597425	0.557459	0.61343	0.737165
F_measure	0.647251	0.636302	0.666819	0.736684
MCC	0.537311	0.48524	0.552584	0.692765
NPV	0.895187	0.915582	0.914497	0.955977
FPR	0.10461	0.084191	0.085277	0.043815
FNR	0.293856	0.258878	0.269614	0.263797

4.13 Comparison of the Proposed Model Using Different Fuzzy Membership Function Generators

The efficacy of the suggested consumer behavior categorization model using three different kinds of fuzzy membership function generators, triangular, Gaussian, and trapezoidal, is examined in Table 8. The triangle membership function exhibits the highest accuracy (0.9309), sensitivity (0.7578), specificity (0.9598), and precision (0.7589) among the three, and consistently produces the most effective results across all performance parameters. Additionally, it has the greatest Matthews Correlation Coefficient (MCC) of 0.7181 and F-measure (0.7584), demonstrating a good balance between true positive and true negative predictions. It is also the most dependable choice for fuzzy rule-based categorization in this situation, with the lowest false positive rate (0.0402) and false negative rate (0.2422). The Gaussian function performs slightly better than trapezoidal, but both are outperformed by the triangular function, underscoring its effectiveness in capturing subtle variations in consumer behavior data.

Table 8.
Proposed Model Using Different Fuzzy Membership Function Generators.

Proposed With a Proposed With a Proposed

Trapezoidal Membership Gaussian Fuzzy (Triangular Membership

Function Generator Membership Function Function Generator)

FPR 0.044722 0.042002 0.040189

NPV 0.955078 0.957775 0.959573

MCC 0.686442 0.705411 0.718057

Accuracy 0.92317 0.927814 0.93091

Sensitivity 0.730803 0.747002 0.757801

F_measure 0.731263 0.747524 0.758364

FNR 0.269197 0.252998 0.242199

Precision 0.731724 0.748047 0.758929

Specificity 0.955278 0.957998 0.959811

	Proposed With a	Proposed With a	Proposed
FPR	0.044722	0.042002	0.040189
NPV	0.955078	0.957775	0.959573
MCC	0.686442	0.705411	0.718057
Accuracy	0.92317	0.927814	0.93091
Sensitivity	0.730803	0.747002	0.757801
F_measure	0.731263	0.747524	0.758364
FNR	0.269197	0.252998	0.242199
Precision	0.731724	0.748047	0.758929
Specificity	0.955278	0.957998	0.959811

5. Discussion

According to MCC, it exhibits exceptional sensitivity, specificity, precision, and overall balanced performance. Accuracy values for these models (LINKNET, SQUEEZENET, LSTM, SVM, CNN, DCNN, BiGRU, RNN, and DNN) range from about 80.7% to 86.5%. This suggests that using the information retrieved during the preprocessing and feature extraction stages, they do a respectable job of forecasting consumer behavior. Likewise, there is an increase in the F-measure and recall reminder, accordingly. The testing results provide more measurement for the portions mentioned above. The results of the trial show that a very high degree of accurate mobile commercial behavior is achieved by the recommended device architecture. The suggested model's high accuracy and well-balanced metrics make it ideal for real-world uses where accuracy and dependability are essential, like customer behavior research in marketing.

The proposed IFCBMC framework provides valuable implications for managers by enabling more accurate and scalable analysis of consumer behavior in big data environments. Its application supports improved customer segmentation, allowing businesses to tailor marketing strategies and personalize services more effectively. The model's robust classification capabilities also enhance fraud detection, particularly in e-commerce, by identifying abnormal purchasing patterns with greater precision. Additionally, its integration with distributed processing ensures real-time insights, making it suitable for high-traffic platforms where timely decision-making is critical.

6. Conclusion

This study suggested a way to assess consumer behavior using fuzzy rule-based categorization with data mapping and bagging methodologies in a map reduction scheme. At first, the input big data of the consumer behavior were preprocessed via data cleaning, and a modified normalization process was carried out. After that, training and testing datasets were created from the normalized data. Using the MapReduce framework, the training dataset's features such as Correlation and Improved entropy, were extracted from the normalized dataset. Next, an enhanced fuzzy rule-based classification model was used to use the extracted features as input into the various bagging techniques. Similarly, the mapper function received the testing normalized dataset as input. The final classification set is obtained by combining the testing dataset of the mapping function and the training dataset of the bagging procedure. Moreover, these classifiers were shuffled to form the prediction outcome by using the Reducer phase. In terms of the maximum statistical metric, the IFCBMC approach achieved the highest accuracy of 0.960, while traditional approaches, such as LSTM with 0.862, SVM with 0.863, DCNN with 0.860, Bi-GRU with 0.854, RNN [24] with 0.861, and DNN [28], had lower accuracy rates. The use of MapReduce to process huge amounts of data improves scalability but may complicate deployment and maintenance, particularly in terms of cluster management and resource allocation. Fuzzy logic allows for greater flexibility in dealing with uncertainties, yet fuzzy rule-based models may be difficult to grasp.

Develop methodologies for real-time or near-real-time customer behavior analytics, extending beyond the batch processing capabilities of traditional MapReduce frameworks. This could involve the adoption of stream processing architectures and adaptive learning algorithms. The real-time applications are considered in the future work.

Footnotes

List of abbreviations

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

Abedin

M. Z.

Hajek

Sharif

Satu

M. S.

Khan

M. I.

(2023). Modelling bank customer behaviour using feature engineering and classification techniques. Research in International Business and Finance, 65, 101913. https://doi.org/10.1016/j.ribaf.2023.101913

Anand

Vessal

S. R.

Rathi

Ameen

(2021). Show me your mobile and I will tell you who you are: Forecasting consumer compassion and altruism behaviour through smartphone type and usage. Journal of Retailing and Consumer Services, 63. https://doi.org/10.1016/j.jretconser.2021.102657

Argyris

Y. A.

Wang

Kim

Yin

(2020). The effects of visual congruence on increasing consumers’ brand engagement: An empirical investigation of influencer marketing on Instagram using deep-learning algorithms for automatic image classification. Computers in Human Behavior, 112, 106443. https://doi.org/10.1016/j.chb.2020.106443

Bilal

Marjani

Lali

M. I.

Malik

Gani

Hashem

I. A. T.

(2020). Profiling users’ behavior, and identifying important features of review “helpfulness”. IEEE Access, 8, 77227–77244. https://doi.org/10.1109/ACCESS.2020.2989463

Chaudhary

Alam

Al–Rakhami

M. S.

Gumaei

(2021). Machine learning-based mathematical modelling for prediction of social media consumer behavior using big data analytics. Journal of Big Data, 8(73). https://doi.org/10.1186/s40537-021-00466-2

del Rio

Lopez

Benitez

J. M.

Herrera

(2015). A MapReduce approach to address big data classification problems based on the fusion of linguistic fuzzy rules. International Journal of Computational Intelligence Systems, 8(3), 422–437. https://doi.org/10.1080/18756891.2015.1017377

Dengler

Prüfer

(2021). Consumers’ privacy choices in the era of big data. Games and Economic Behavior, 130, 499–520. https://doi.org/10.1016/j.geb.2021.09.007

Horng

S. J.

Chen

Y. H.

Run

R. S.

Chen

R. J.

Lai

J. L.

Sentosal

K. O.

(2009). An improved score level fusion in multimodal biometricsystems. In 2009 international conference on parallel and distributedcomputing, applications and technologies (pp. 239–246). IEEE.

Hung

N. D.

Van-Nam

(2023). Integrated preference argumentation and applications in consumer behaviour analyses. International Journal of Approximate Reasoning, 159. https://doi.org/10.1016/j.ijar.2023.108938

10.

Jeong

(2021). Identifying consumer preferences from user-generated content on Amazon.Com by leveraging machine learning. IEEE Access, 9, 147357–147396. https://doi.org/10.1109/ACCESS.2021.3123301

11.

Yuan

Tian

(2023). Influence of online E-commerce interaction on consumer satisfaction based on big data algorithm. Heliyon, 9, e18322. https://doi.org/10.1016/j.heliyon.2023.e18322

12.

Zhao

(2020). Measuring ease of use of mobile applications in e-commerce retailing from the perspective of consumer online shopping behaviour patterns. Journal of Retailing and Consumer Services, 55, 102093. https://doi.org/10.1016/j.jretconser.2020.102093

13.

Lin

Xiao

(2022). Using social media big data for tourist demand forecasting: A new machine learning analytical approach. Journal of Digital Economy, 1, 32–43. https://doi.org/10.1016/j.jdec.2022.08.006

14.

Meena

Kumar

(2022). Online food delivery companies’ performance and consumers expectations during Covid-19: An investigation using machine learning approach. Journal of Retailing and Consumer Services, 68, 103052. https://doi.org/10.1016/j.jretconser.2022.103052

15.

Mikalef

Sharma

Chatterjee

Chaudhuri

Parida

Gupta

(2023). All eyes on me: Predicting consumer intentions on social commerce platforms using eye-tracking data and ensemble learning. Decision Support Systems, 175, 114039. https://doi.org/10.1016/j.dss.2023.114039

16.

Mirashk

Albadvi

Kargari

Javide

Eshghi

Shahidi

(2019). Using RNN to predict customer behavior in high volume transactional data. Springer Nature Switzerland AG.

17.

Mou

Benyoucef

(2021). Consumer behavior in social commerce: Results from a meta-analysis. Technological Forecasting & Social Change, 167, 120734. https://doi.org/10.1016/j.techfore.2021.120734

18.

Navin

Krishnan

(2024). Fuzzy rule based classifier model for evidence based clinical decision support systems. Intelligent Systems with Applications, 22, 200393. https://doi.org/10.1016/j.iswa.2024.200393

19.

Prasetyo

Setiawan

N. A.

Adji

T. B.

(2020). Improving normalization method of higher-order neural network in the forecasting of oil production. E3S Web of Conferences, 200, 02016. https://doi.org/10.1051/e3sconf/202020002016

20.

Pratap

(2022). Consumer buying behavior analysis dataset, Kaggle. https://www.kaggle.com/code/aradhanapratap/consumer-buying-behavioranalysis/notebook

21.

Raj Kannan

Sabitha

Karthik

Shanthini

(2021). Mouse movement pattern based analysis of customer behavior (CBA-MMP) using cloud data analytics. Wireless Personal Communications, 117, 2885–2901. https://doi.org/10.1007/s11277-020-07055-1

22.

Reyes-Menendez

Palos-Sanchez

Ramon Saura

Rodríguez Santos

(2022). Revisiting the impact of perceived social value on consumer behavior toward luxury brands. European Management Journal, 40(2), 224–233. https://doi.org/10.1016/j.emj.2021.06.006

23.

Sharifi

Shokouhyar

(2021). Promoting consumer’s attitude toward refurbished mobile phones: A social media analytics approach. Resources, Conservation and Recycling, 167, 105398. https://doi.org/10.1016/j.resconrec.2021.105398

24.

Shiu-Li

Yi-Hsien

(2022). Exploring consumer online purchase and search behavior: An FCB grid perspective. Asia Pacific Management Review, 27, 245–256. https://doi.org/10.1016/j.apmrv.2021.10.003

25.

Simona-Vasilica

Bara

Tudorica

B. G.

Alinoiu

M. I.

Botezatu

M. A.

(2021). Insights into demand-side management with big data analytics in electricity consumers’ behaviour. Computers and Electrical Engineering, 89, 106902. https://doi.org/10.1016/j.compeleceng.2020.106902

26.

Sitharamulu

Prasad

K. R.

Reddy

K. S.

Prasad

A. K.

Dass

M. V.

(2024). Hybrid classifier model for big data by leveraging map reduce framework. International Journal of Data Mining, Modelling and Management, 16(1), 23–48. https://doi.org/10.1504/IJDMMM.2024.136219

27.

Stangherlin

I. C.

Thøgersen

Dutra de Barcellos

(2023). Behavioral spillover in the circular economy: The importance of consumer goals. Journal of Environmental Psychology, 91, 102123. https://doi.org/10.1016/j.jenvp.2023.102123

28.

Sundararaj

Rejeesh

M. R.

(2021). A detailed behavioral analysis on consumer and customer changing behavior with respect to social networking sites. Journal of Retailing and Consumer Services, 58, 102190. https://doi.org/10.1016/j.jretconser.2020.102190

29.

Taghikhah

Voinov

Shukla

Filatova

(2021). Shifts in consumer behavior towards organic products: Theory-driven data analytics. Journal of Retailing and Consumer Services, 61, 102516. https://doi.org/10.1016/j.jretconser.2021.102516

30.

Tao

Zhou

(2023). Can online consumer reviews signal restaurant closure: A deep learning-based time-series analysis. IEEE Transactions on Engineering Management, 70(3), 834–848. https://doi.org/10.1109/TEM.2020.3016329

31.

Tassell

Aurisicchio

(2023). Refill at home for fast-moving consumer goods: Uncovering compliant and divergent consumer behaviour. Sustainable Production and Consumption, 39, 63–78. https://doi.org/10.1016/j.spc.2023.04.018

32.

Trawinski

Cordon

Quirin

(2011). On designing fuzzy rule-based multiclassification systems by combining FURIA with bagging and feature selection. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 19(4), 589–633. https://doi.org/10.1142/S0218488511007155

33.

Trovati

Zhang

Ray

(2020). An entropy based approach to real-time information extraction for industry 4.0. IEEE Transactions on Industrial Informatics, 16(9), 6033–6041. https://doi.org/10.1109/TII.2019.2962029

34.

Ulph

Panzone

Hilton

(2023). Do rational people sometimes act irrationally? A dynamic self-regulation model of sustainable consumer behavior. Economic Modelling, 126, 106384. https://doi.org/10.1016/j.econmod.2023.106384

35.

Xue

Caiguo

Chenxia

(2022). Consumption of traditional handicraft fashion: Motivations, intentions and behaviours of Chinese consumers. Cleaner and Responsible Consumption, 4, 100046. https://doi.org/10.1016/j.clrc.2021.100046

36.

Yuan

(2022). Analysis of consumer behavior data based on deep neural network model. Journal of Function Spaces, 2022, 1–10. https://doi.org/10.1155/2022/4938278

37.

Zaremohzzabieh

Ismail

Ahrari

Abu Samah

(2021). The effects of consumer attitude on green purchase intention: A meta-analytic path analysis. Journal of Business Research, 132, 732–743. https://doi.org/10.1016/j.jbusres.2020.10.053

38.

Zhao

Yang

Fei

Zhai

(2019). An improved belief entropy to measure uncertainty of basic probability assignments based on Deng entropy and belief interval. Entropy, 21, 1122. https://doi.org/10.3390/e21111122

	Model Without	Model With Conventional	Model With Conventional
Metrics	Feature Extraction	Normalization	Entropy	IFCBMC
Accuracy	0.782	0.812	0.857	0.931
NPV	0.835	0.869	0.917	0.960
FPR	0.164	0.130	0.083	0.040
Sensitivity	0.681	0.703	0.752	0.758
F-measure	0.463	0.468	0.501	0.758
FNR	0.319	0.297	0.248	0.242
Specificity	0.836	0.870	0.917	0.960
MCC	0.388	0.387	0.418	0.718
Precision	0.463	0.468	0.501	0.759

	Proposed	Proposed With	Proposed With	Proposed	Proposed With
	Without	z-score_	Robust_	With	Conventional_
Metrics	Normalization	Normalization	Normalization	l1_Normalization	Normalization	Proposed
MCC	0.390037	0.395609	0.391894	0.389503	0.387111	0.718057
Specificity	0.840589	0.85585	0.845676	0.857656	0.869635	0.959811
FPR	0.159411	0.14415	0.154324	0.142344	0.130365	0.040189
F_measure	0.465177	0.472364	0.467572	0.467578	0.467584	0.758364
Sensitivity	0.685884	0.699079	0.690282	0.696627	0.702972	0.757801
Precision	0.465523	0.472716	0.46792	0.467926	0.467932	0.758929
NPV	0.840381	0.855638	0.845467	0.857443	0.86942	0.959573
Accuracy	0.78683	0.800935	0.791531	0.801798	0.812064	0.93091
FNR	0.314116	0.300921	0.309718	0.303373	0.297028	0.242199

	Proposed With	Proposed With	Proposed With	Proposed With Only
	Only Shannon	Only Belief	Only Deng	Improved Deng
Metrics	Entropy	Entropy	Entropy	Entropy Features
Accuracy	0.852649	0.864421	0.871487	0.924718
Sensitivity	0.706144	0.741122	0.730386	0.736203
Specificity	0.89539	0.915809	0.914723	0.956185
Precision	0.597425	0.557459	0.61343	0.737165
F_measure	0.647251	0.636302	0.666819	0.736684
MCC	0.537311	0.48524	0.552584	0.692765
NPV	0.895187	0.915582	0.914497	0.955977
FPR	0.10461	0.084191	0.085277	0.043815
FNR	0.293856	0.258878	0.269614	0.263797

	Proposed With a	Proposed With a	Proposed
	Trapezoidal Membership	Gaussian Fuzzy	(Triangular Membership
	Function Generator	Membership Function	Function Generator)
FPR	0.044722	0.042002	0.040189
NPV	0.955078	0.957775	0.959573
MCC	0.686442	0.705411	0.718057
Accuracy	0.92317	0.927814	0.93091
Sensitivity	0.730803	0.747002	0.757801
F_measure	0.731263	0.747524	0.758364
FNR	0.269197	0.252998	0.242199
Precision	0.731724	0.748047	0.758929
Specificity	0.955278	0.957998	0.959811

A Fuzzy Rule-Based Framework for Consumer Behavior Analytics

Abstract

Keywords

1. Introduction

1.1 Research Objectives

1.2 Research Questions

2. Literature Review

3. Proposed Framework for Consumer Behavior Analytics via Fuzzy Rule-Based Classification

3.1.1 Data Cleaning

3.1.2 Modified Normalization

3.2.1.1 Improved Entropy F E

4.1 Experimental Setup

4.2 Dataset Description

4.3 Performance Analysis

4.4 Comparative Study on Positive Metrics

6. Conclusion

Footnotes

List of abbreviations

Funding

Declaration of Conflicting Interests

References

3.2.1.1 Improved Entropy $F_{E}$