Empowering Retail in the Metaverse by Leveraging Consumer Behavior Analysis for Personalized Shopping: A Pilot Study in the Saudi Market

Abstract

The integration of advanced technologies, such as the Metaverse, has the potential to revolutionize the retail industry and enhance the shopping experience. Understanding consumer behavior and leveraging machine learning predictions based on analysis can significantly enhance user experiences, enabling personalized interactions and fostering overall engagement within the virtual environment. In our ongoing research effort, we have developed a consumer behavior framework to predict interesting buying patterns based on analyzing sales transaction records using association rule learning techniques aiming at improving sales parameters for retailers. In this paper, we introduce a validation analysis of our predictive framework that can improve the personalization of the shopping experience in virtual reality shopping environments, which provides powerful marketing facilities, unlike real-time shopping. The findings of this work provide a promising outcome in terms of achieving satisfactory prediction accuracy in a focused pilot study conducted in association with a prominent retailer in Saudi Arabia. Such results can be employed to empower the personalization of the shopping experience, especially on virtual platforms such as the Metaverse, which is expected to play a revolutionary role in future businesses and other life activities. Shopping in the Metaverse offers a unique blend of immersive experiences and endless possibilities, enabling consumers to interact with products and brands in a virtual environment like never before. This integration of cutting-edge technology not only transforms the retail landscape but also paves the way for a new era of personalized and engaging shopping experiences. Lastly, this empowerment offers new opportunities for retailers and streamlines the process of engaging with customers in innovative ways.

1. Introduction

The term Metaverse has been interpreted and used in diverse ways. In 2004, Ref. [1] described it as combining an online and real-world environment where users can engage and socialize with others, conduct business, and enjoy themselves, utilizing the real world as a metaphor. In 2013, Ref. [2] defined it as a computer-generated universe beyond the real world, a fully immersive 3D digital environment that represents the entirety of shared online space across all representational dimensions. Ref. [2] conceptualized the term ‘Metaverse’ as a combination of the prefix ‘meta’ (meaning ‘beyond’) and the suffix ‘verse’ (meaning ‘universe’). Currently, the rapid development of technology in the fourth Industrial Revolution primarily supports Metaverse development, industries, and societal environments and processes due to the expansion of novel technologies like blockchain, artificial intelligence (AI), augmented reality (AR), virtual reality (VR), and enhanced mobile networks [3,4,5].

In e-commerce, the Metaverse is one of the latest developments. The Metaverse allows retailers to create virtual environments where consumers can explore products in a 3D space, enhancing the shopping experience beyond traditional online or physical stores [6,7,8]. This immersive experience can lead to increased engagement and satisfaction among consumers [6,9]. In the realm of the Metaverse and virtual reality, the growing significance of its influence on the business world is becoming increasingly pronounced [6]. The supermarket in the Metaverse, a relatively new development in online retail, allows users to experience a sense of physical presence in a digital shopping environment, effectively bridging the gap between physical and virtual realms [7]. With ongoing technological advancements, consumers are increasingly embracing product exploration and decision-making. In business marketing, the key to success lies in effectively influencing consumer behavior, enabling companies to tailor their strategies and enhance customer satisfaction levels [8,10].

The integration of virtual reality (VR) and the emerging concept of the Metaverse is reshaping the retail landscape, offering unprecedented opportunities for retailers to engage with customers in immersive digital environments [8,11,12]. In the realm of supermarkets within the Metaverse, this technological convergence presents a paradigm shift in how consumers interact with products and brands. The virtual supermarket is a new concept, referring to a digital or online shopping experience where customers can virtually navigate through a supermarket’s aisles and shelves using a computer or mobile device [13,14]. Virtual supermarkets leverage advanced technologies such as VR headsets and haptic feedback systems to create lifelike shopping experiences where users can browse aisles, interact with products, and make purchases in a virtual setting. According to [8,12,15], VR simulations in retail settings have been shown to significantly enhance user engagement and satisfaction, leading to increased purchase intent and brand loyalty. Moreover, as highlighted in [8,12,16], the Metaverse offers retailers a unique platform to experiment with innovative strategies such as personalized product recommendations based on user behaviors, dynamic pricing adjustments, and interactive shopping experiences. In the context of supermarkets, association rule learning techniques, as mentioned by Refs. [12,17], can be applied to analyze shopping patterns and optimize product placements, ultimately improving the efficiency of virtual store layouts and enhancing the overall shopping journey for customers in the Metaverse.

Within the framework of Metaverse shopping considerations, analyzing consumer behavior provides valuable insights into preferences, decision-making processes, and potential adoption barriers [6,8,12,18]. The Metaverse enables the development of precise marketing strategies through detailed user data analysis in trade. By harnessing advanced analytics and machine learning, businesses can create personalized campaigns based on individual preferences. Immersive technologies like virtual reality enhance engagement through interactive experiences, fostering strong brand connections. Ultimately, the Metaverse transforms marketing by offering tailored, engaging, and adaptable strategies that resonate deeply with consumers [17,19]. Understanding the intricacies of consumer behavior, especially prior to the transition to virtual environments, is crucial for companies aiming to capitalize on the growing marketing trends in virtual environments [8,17,20].

There are many enablers merged with the Metaverse as shown in Figure 1, and those enablers provide retailers with tools to create immersive and personalized shopping experiences for customers [21,22,23]. Technologies like virtual reality (VR) and augmented reality (AR) enable retailers to engage customers in novel ways [7,24,25]. Moreover, VR/AR allows users to immerse themselves in virtual worlds or overlay digital information onto the real world, providing the foundation for realistic and interactive Metaverse experiences. Blockchain technology can enable secure and transparent transactions within the Metaverse, allowing for the creation of digital assets and virtual economies [3,26,27]. Also, high-speed internet connectivity is an enabler that provides fast and reliable internet connections that are essential for seamless Metaverse experiences, allowing for smooth interactions and real-time data transfer [28,29]. Lastly, artificial intelligence (AI) can power many aspects of the Metaverse, from generating realistic virtual environments to enabling intelligent virtual agents and personalized experiences [27,30,31]. AI is a cornerstone of the Metaverse, powering personalized experiences, intelligent virtual agents, and data-driven insights. It enhances user interactions, optimizes virtual environments, and drives innovation in the Metaverse ecosystem [25,27]. The existence of Metaverse enablers empowers retailers to adapt to changing consumer preferences, offer innovative shopping experiences, optimize operations, and create new opportunities for growth in the digital realm. By leveraging these technologies effectively, retailers can thrive in the dynamic landscape of virtual retail.

The emergence of the Metaverse, bolstered by advancements in AI and other technologies, presents a unique opportunity to revolutionize the shopping experience in virtual supermarkets [25,27,30]. AI, as a pivotal enabler within the Metaverse, offers capabilities for analyzing vast amounts of data, understanding user preferences, and delivering tailored recommendations [25]. By harnessing AI, virtual supermarkets can provide personalized shopping experiences that cater to individual tastes and needs, mirroring the level of customization found in physical retail settings. This integration of AI in the metaverse not only streamlines decision-making processes for consumers but also enhances the overall shopping journey by offering targeted suggestions and creating a more engaging and interactive environment. However, there remains a significant gap in the current literature regarding the practical implementation and validation of these AI-driven techniques within the metaverse to truly enhance the personalized shopping experience [32]. Bridging this gap is crucial to demonstrate the real-world impact and effectiveness of such innovations in transforming the way consumers shop within virtual environments.

Consumer behavior in the digital age is increasingly shaped by personalization and customization preferences [8,12,33]. The Metaverse, with its immersive and interactive nature, has the potential to markedly improve personalized shopping experiences by aligning with these evolving consumer expectations [12,34]. Through virtual supermarkets in the Metaverse, shoppers can explore products in a dynamic and personalized manner, receiving recommendations based on their preferences and past interactions [32,35]. This tailored approach not only fosters a sense of individuality and uniqueness but also enhances consumer satisfaction and loyalty.

In our previous work [17], we proposed a multi-layered comprehensive framework to identify trends in consumer behavior using multiple layers of processing, including clustering, classification, and association rule learning. The aim was to help retailers understand customer behavior by utilizing the power of big data analysis and techniques. The philosophy behind building a consumer behavior analysis framework as a system of layers is to provide a standardized and consistent set of tools and methods for analyzing consumer behavior from different data sources. Figure 2 shows the framework as a layered system and the specific technique used in each layer. The bottom layer, Layer 1 (Raw Data Layer): in this layer, we used two data sources, the loyalty program repository and the data of sales transactions. Layer 2 (Data Preprocessing Layer): where we ensure that the data are cleaned and normalized, and any missing or incorrect data are removed. Layer 3 (Processing and Perception Layer): this is the backbone of the framework where the technical part is applied. In this layer, we implemented two machine learning techniques: an unsupervised data clustering technique using a K-means algorithm, and association rule learning using an FP-Growth algorithm. Layer 4 (Marketing Intelligence layer): a post-processing stage that supports the understanding of all results and produces more valuable findings. Here, we used a THAPE: Tunable Hybrid Associative Predictive Engine, which is a novel approach for enhancing the interpretability in association rule learning for retail [36]. Layer 5 (Application Layer): the last layer, where the results and recommendations are applied. For more details on the framework’s layers and the THAPE approach refer to [17,36].

The THAPE represents a significant advancement in enhancing the personalized shopping experience within virtual supermarkets, serving as a key enabler in the Metaverse environment. Acting as a crucial feeding engine, the THAPE bridges the gap between the physical world and the Metaverse by providing real-time predictions and valuable insights. This innovative engine integrates various AI techniques, including big data analysis, association rule learning, clustering, and classification, to drive its predictive capabilities and optimize the customization of user experiences.

The main objective of this research is to support and enhance the shopping experience in the Metaverse by leveraging association rules and prediction of customers’ purchase desires. This approach serves as an enabler for the Metaverse, providing a customized and distinctive supermarket shopping experience for each individual in a virtual reality supermarket, as rearrangements and product suggestions become much easier. Such enhancements improve the personalized shopping experience on virtual reality platforms, enabling consumers to fully embrace and enjoy this innovative way of interacting.

To sum up, this paper’s contribution can be summarized in three points:

• Presenting a comprehensive review of the Metaverse in the retail sector and its enablers in that context.
• Proposing the THAPE as an enabler for personalized shopping in the Metaverse environment.
• Evaluating THAPE accuracy and benefits based on a case study experiment in association with a major retailer in Saudi Arabia.

The paper is structured as follows: Section 2 reviews related work and the current state of virtual shopping as well as the challenges in virtual shopping. Section 3 presents the previous analysis, while Section 4 explores the business perspective. Section 5 discusses the Metaverse as a platform for retailers. Section 6 introduces the proposed architecture, the THAPE, as an enabler for personalized shopping in the Metaverse. Section 7 focuses on consumer behavior prediction, followed by Section 8, which presents the results and analysis. Section 9 discusses future research directions, and finally, Section 10 concludes the paper.

2. Related Work

2.1. Metaverse and Retail

The Metaverse has evolved significantly from its early days as a media novelty to a transformative technology utilized across various sectors, including retail [7,14,37]. This section offers a comprehensive overview of the current state of the Metaverse in supermarket shopping. It covers various aspects, including technological advancements and their challenges, consumer behavior, and the implications of the Metaverse in retail.

Researchers in [24,38] provided a multidisciplinary perspective on the Metaverse and augmented virtuality (AV), focusing on its potential implications for consumer behavior, particularly in the domains of consumer identity, social influence, and ownership. One study explored the impact of health-targeting nutrition apps on shopping behavior using a Metaverse supermarket simulation and a virtual replica smartphone with augmented virtuality (AV) features. The study presented a hybrid concept, where the camera input was extended with WLAN-transmitted data to enhance tracking and interaction with self-made applications. Reference [24] proposed the process of integrating real smartphones via AV and described solutions to usability challenges faced during the development of a virtual supermarket. Moreover, features for a virtual smartphone replica to interact with virtual nutrition products in a virtual environment have been explored.

Reference [38] investigated the influence of perceived consumer experience, perceived brand engagement, and gamification on the intention to use the Metaverse of a specific brand. The study found that all three variables had a positive and significant relationship with the intention to use. This study extended knowledge in the technology acceptance space, introducing three new factors influencing adoption within the TAM framework: perceived brand engagement, perceived consumer experience, and gamification. The study also highlighted that companies must consider these factors when implementing the Metaverse in business to ensure higher utilization.

Some researchers have also proposed architecture for virtual agents to grasp social practices in the Metaverse. Reference [38] proposed an architecture for designing and developing virtual agents that can understand social practices in the Metaverse and act accordingly, respecting social norms and exhibiting credible and acceptable social behavior. The proposed architecture provided a robust and efficient infrastructure for developing social NPCs in the Metaverse, enabling them to adapt their behavior according to the current situation in a supermarket scenario. The methods employed included developing a virtual environment using Unity, integrating Metaverse-specific SDKs, and utilizing deep learning modules for object recognition, expression recognition, and gesture recognition. The Scene Analysis layer employed EfficientNET and YOLOv4 for object recognition, Emotion FERplus for emotion recognition, and a CNN for handwritten digit recognition. The NLP component incorporated transformer models such as BERT and GPT-3. Additionally, 3D modeling was used with MakeHuman software, along with the creation of a virtual scene in Unity.

2.2. Metaverse Enablers

The Metaverse has many enablers, ranging from hardware and equipment to networking, computing, virtual platforms, and artificial intelligence (AI) technologies. This section presents related work in the field of AI technologies [14,37,39] and discusses the broader technological enablers shaping the Metaverse.

Future trends in the Metaverse are increasingly influenced by IoT, blockchain, and AI techniques [27,40]. Reference [27] provided an overview of the technology roadmap for the Metaverse, discussing the roles of IoT, blockchain, AI, and other technologies in shaping its future across various domains, with a specific focus on the medical field. The study highlighted the role of IoT devices in providing traceable data identified by unique tags, used in a blockchain-based Metaverse. However, the study did not address the limitations of the Metaverse in this context.

Reference [40] proposed a methodology for developing embodied conversational agents capable of interacting as conversational meta bots in virtual worlds using AI techniques. These meta bots’ conversational behavior was based on a statistical dialog model trained on a corpus of dialogs. Reference [41] discussed aspects of social influence, limitations, and open challenges in the Metaverse. The study emphasized that the Metaverse can provide a tangible sensation, allowing users to interact with virtual objects and environments in a more immersive and interactive way. However, it also identified challenges such as the need for improvements in hardware and software, as well as the potential for addiction and social influence.

Table 1. AI application summary with highlights to the method, datasets, objective, and outcomes.

Category	Description Summary	Key Components
Framework	ALMAA framework (adaptive learning model for ai agents). Using detailed user interaction logs, the framework shows good results in understanding and analyzing adaptive AI in virtual settings of Epic Games and AltspaceVR [42].	AI agents
	A deep learning framework with multiple CNNs for feature extraction and prediction, used to forecast eye fixations in a task-oriented virtual environment in virtual reality. Using users’ eye tracking data, FixationNet outperformed the state-of-the-art method (DGaze) by 19.8%. Strong correlations were found between eye fixations and factors like historical gaze positions, task-related objects, saliency information, and head rotation velocities [43].	CNNs
Recommendation system	The recommendation system is built using a Neural Collaborative Filtering (NCF) model, which generates recommendations based on purchase history and user interaction data in Amazon [16].	NCF + CNN
	Build a realistic and practical metaverse by pushing the frontier of fundamental technologies, including perception, computation, reconstruction, cooperation, and interaction [44].	Cloud computing
	Virtual shopping mall system with different sensors (proximity sensor, visibility sensor, and touch Sensor) used to track the customer’s avatar and provide personalized product recommendations from real-life data, including sales transactions [30].	Sensors

summarizes AI applications in the Metaverse, highlighting methods, datasets, objectives, and outcomes. Reference [42] proposed an adaptive learning model for AI agents (ALMAA) as a framework for understanding and analyzing adaptive AI in virtual settings. The study investigated adaptive learning mechanisms in AI agents within the Metaverse and their transformative potential in dynamic virtual environments. The results demonstrated that the ALMAA enhanced user experience and operational efficiency, with AI agents showing improved response times and decision accuracy.

In the same field of AI, certain artificial intelligence algorithms have been integrated into virtual reality hardware to enhance human–computer interactions using visual data. A deep learning architecture incorporating multiple convolutional neural networks was developed by [43] to anticipate the user gaze in gaze-dependent applications like content design and rendering. This framework processes virtual reality images, gaze-tracking data, and head movement data, improving interactive experiences in the Metaverse.

Reference [44] re-introduced the Metaverse within a new framework that combined it with AI. This new framework was built on a broad range of technologies, including perception, computation, reconstruction, cooperation, and interaction technologies. As highlighted in [44], significant progress has been made in artificial intelligence-powered technologies, which are essential for building a realistic and practical Metaverse.

The VR supermarket platform proposed in Reference [16] provides a realistic and user-friendly shopping experience, allowing users to explore and buy products in a 360° rendered showroom. The integrated recommendation system provides personalized product suggestions based on purchase history and user interaction. The proposed system, using NCF + CNN + Attention, is evaluated on the Amazon dataset and shows promising results in recommending products mitigating the new user problem and sparsity problem, making it suitable for a VR supermarket application [16]. The proposed model’s validation MAE is 0.6188, indicating that it can generalize to new data and mitigate the new user problem and sparsity problem.

In [30], the author demonstrated the effectiveness of the proposed approach in providing personalized product recommendations in a 3D virtual shopping mall. Reference [30] integrated virtual reality and data mining approaches to realize the recommendation of products and life-like navigation in 3D virtual shopping malls. The author in [45] used a laboratory experiment with participants randomly assigned to either a high-immersive VR shopping environment or a low-immersive desktop environment. The study used a two-stage approach, first confirming the reliability of the measurement model and then analyzing the structural model, including testing the hypotheses. The experiment involved a conjoint-analytic design, with participants making purchasing decisions in a VR shopping environment.

There are many fields that merged with the Metaverse such as virtual reality (VR), augmented reality (AR), and immersive/mixed reality.

Table 2. Summary with highlights to fields, methods, datasets, objectives, and outcomes.

Field	Method	Objective	Outcomes	Year
Virtual reality + Collected sales transaction data [46]	Using the retail environment setup component to present an example of VR shopping applications developed with VRSI.	To help in abstracting the layers needed to set up VR technology on applications developed with Unity, and to provide data registration tools that monitor user activity from non-invasive data sources.	An example of a VR shopping application developed with VRSI was presented, and the public repository with all the data and source codes was made fully open.	2024
Virtual reality + Real experiment [47]	Employed a 2 (Interactivity: high vs. low) × 2 (Vividness: high vs. low) × 2 (Product type: leather chair vs. flower) between-subjects experiment design, with eight VR store versions created to correspond to eight scenarios. Each participant was randomly assigned to experience a VR store scenario.	Investigated how virtual system characteristics impacted consumers’ impulsive buying decisions in VR-based online shopping and contributed to the growing body of e-commerce studies on online impulse buying and the effectiveness of VR systems on consumers’ experiences and consequent behavioral decisions.	The path coefficients of the six proposed hypotheses related to the stimuli–organism route were statistically significant, indicating that VR shopping platforms evoke both individual cognitive and affective reactions. Additionally, impulsiveness positively influenced the urge to buy impulsively.	2023
Immersive/mixed reality + MNIST dataset and Unity data [38]	Deep learning modules for all recognitions. Unity for the development of the virtual environment. NLP component incorporated transformer modules such as BERT, GPT-3, and Tortoise-TTS.	To develop a social practice-based architecture for social NPCs in the Metaverse, enabling them to adapt their behavior according to the current situation in a supermarket scenario.	Enabled the agent to adapt its behavior according to the current situation in a supermarket scenario, recognizing objects, emotions, and gestures, and using NLP to fine-tune pre-trained models.	2022
Virtual reality + Online data collected [38]	Collecting data through a structured survey utilizing the Likert scale with five response items. The data analysis was carried out using Structural Equation Modeling (SEM) in determining the causal relationship between latent variables and the dependent variable.	To identify key factors influencing consumers to adopt metaverse technology, specifically examining the roles of perceived consumer experience, perceived brand engagement, and gamification.	The results showed that the perceived consumer experience had positive and significant effects on the intention to use Metaverse technology of a specific brand. The results showed that all indicators of Perceived Consumer Experience had a loading factor > 0.7, indicating good convergent validity. The path coefficient test showed that the latent variables had good reliability.	2022
Augmented reality + WLAN data [24]	Using virtual supermarket simulation and combining an add-on RGB camera on the augmented reality (AR) headset to track the screen of the participants.	To investigate the influences of health-related nutrition apps on shopping behavior using a virtual supermarket simulation and a virtual replica smartphone with augmented virtuality features.	The study showed high potential in understanding the purchasing decisions of participants and enhancing the effectiveness of health-oriented applications. Also, the study found that the hybrid approach using an RGB camera and fiducial markers allowed for high-quality tracking of the participant’s smartphone screen and integration of real-world apps into the simulation.	2021
Immersive/mixed reality + Laboratory experiments data [45]	The analysis used covariance-based structural equation modeling (CB SEM) and a multivariate analysis of variance (MANOVA) to test the effects of treatment manipulation on the dependent variables.	To develop and experimentally validate a theoretical model that explains how immersion affects adoption of VR shopping environments.	Immersion had a positive effect on perceived telepresence, but a negative effect on perceived product diagnosticity. However, when controlling readability, high immersion had a positive influence on the intention to reuse the shopping environment.	2019

summarizes references from this area, highlighting the fields, methods, datasets, objectives, and outcomes.

The Virtual Reality Shopping Insights (VRSI) framework, introduced by [46], is an artificial intelligence (AI)-based framework designed to facilitate the creation of virtual reality (VR) shopping applications while streamlining the process of data collection and analysis. The VRSI aims to support software developers and researchers by simplifying the integration of VR technology within applications built using Unity. Additionally, this framework incorporates AI concepts to enhance user experience, enabling personalized recommendations and adaptive interfaces based on user behavior. By assisting in the development of VR shopping applications, the VRSI framework also plays a crucial role in gathering and evaluating user data to inform design choices effectively. Paper [48] delved into consumer behavior within the Metaverse and emphasized the impact of real-time predictive analytics, simulation modeling tools, and computer vision algorithms on enhancing immersive retail experiences. The study highlighted the importance of virtual content optimization in engaging Metaverse events, achieved through the integration of retail analytics, social commerce, and deep learning-based ambient sound processing in 3D virtual environments. Reference [48] proved that real-time predictive analytics, simulation modeling tools, and computer vision algorithms play essential roles in shaping immersive retail experiences throughout the customer journey.

There is much recent research on virtual reality shopping, [47] investigating the effectiveness of virtual reality (VR) shopping platform factors on consumers’ internal stages and their impact on impulse buying behavior in virtual shopping stores. Reference [47] used a 2 × 2 × 2 factorial design experiment with 227 participants, and the measurement of this study consisted of seven constructs, namely Interactivity, Vividness, Telepresence, Perceived Diagnosticity, Playfulness, Impulsiveness, and Urge to Buy Impulsively. The study found that both interactivity and vividness positively influenced internal perceptions, which in turn triggered the urge to buy impulsively.

2.3. Challenges in Metaverse Virtual Shopping

Today, the Metaverse enables customers to access virtual shops where they can explore well-presented products and make informed purchase decisions from the comfort of their homes. A critical aspect of this experience is how consumers navigate the digital design of these virtual marketplaces [49]. However, several challenges remain in the realm of VR shopping. One major issue is the absence of physical interaction, which limits the sensory experience that traditional shopping offers [50,51]. Consumers often miss the tactile feedback of handling products, leading to concerns about trust and security, especially in light of rising cybercrime and information theft [52]. This skepticism can discourage users from sharing personal details or making purchases in virtual environments [52,53].

The design of VR shopping platforms is crucial for creating an engaging user experience. Poorly organized interfaces and inconvenient navigation can deter potential customers, preventing them from exploring new products or brands [54,55]. Additionally, social influences, such as peer recommendations, which are significant in physical stores, may not translate effectively to virtual environments.

Despite these challenges, VR technology presents unique opportunities for enhancing e-commerce. It allows for richer, more immersive shopping experiences that can significantly increase consumer engagement and satisfaction. Research indicates that VR can enhance decision-making by enabling users to interact with products in ways that traditional online shopping cannot. This includes simulating product usage and providing personalized shopping experiences tailored to individual preferences.

Nonetheless, the high costs associated with VR hardware and software pose significant barriers to widespread adoption, particularly for small and medium-sized enterprises [56]. Technical challenges related to compatibility, latency, and bandwidth also hinder integration into existing e-commerce platforms [57,58,59]. Furthermore, privacy and data security concerns remain paramount, necessitating robust measures to build consumer trust in VR-based shopping [60].

As VR continues to evolve, it offers retailers the potential to create innovative marketing campaigns and unique shopping experiences [19]. In summary, while VR shopping presents both challenges and opportunities, its ability to reshape the online retail landscape is undeniable. The future of VR in e-commerce holds promise for creating immersive, personalized shopping experiences that meet the needs and expectations of modern consumers.

3. Previous Analysis

In [17], we proposed a comprehensive machine learning-enabled framework for analyzing trends in consumer behavior, gaining insight into the data, and enabling data-driven decision-making. Towards building the framework, in an iterative Spiral model, we analyzed business needs and identified demands and performance indicators with the business owner of the third strongest retailer in the Arab world [53,54]. The framework presented in the Introduction Section shows the techniques used in each layer as a case study of the original framework, utilizing a real dataset retrieved from the business owner, which includes three months of sales transaction records for more than 30,000 customers. Figure 3 summarizes the detailed steps we have selected from the multi-layered framework to predict consumer behavior based on association rule learning. Below is what we have carried out and the output at each stage:

• First, we took the raw data, cleaned it, and prepared it for processing.
• The data from the loyalty program, which includes consumers’ personal information, was processed using a K-means clustering algorithm. The clustering process considered individual personal details, the frequency of monthly visits, and the average basket size.

  ○

  Output: Three classes of consumers, Class A, Class B, and Class C. Notably, the cluster of Class C included the largest number of consumers and had the highest average monthly spending per individual, despite having the lowest average number of visits per month.

• The sales transaction data, which included three months’ worth of transactions, were processed using an association rule learning algorithm (FP-Growth algorithm).

  ○

  Output: A large set of association rules.

• Next, we applied a classification algorithm (random forest classifier) alongside the data dictionary to classify the generated rules.

  ○

  Output: A binary classification indicating whether an association rule is considered interesting or not interesting based on the trained algorithm’s predictions.

• Finally, we ran our tool, the THAPE: Tunable Hybrid Associative Predictive Engine.

  ○

  The THAPE is a post-processing tool designed to enhance the interpretability of association rules because there will still be a large number of interesting generated rules.

  ○

  The resulting rules were filtered in two ways: price-based and distance-based. Price-based filtering prioritizes association rules that involve products with the highest price difference between the left-hand side (LHS) and right-hand side (RHS). These ARs feature a cheaper product in the LHS, such as ‘cat food’, leading to a higher-value product in the RHS, such as a ‘microwave oven’, while distance-based filtering prioritizes association rules based on the maximum physical distance between the left-hand side (LHS) and the right-hand side (RHS). These ARs feature the farthest product in the LHS, such as ‘Shampoo,’ leading to a cross-product in the RHS, such as ‘Milk’.

  ○

  After filtering, the results were sorted in descending order based on confidence level. Finally, we selected the top ten rules for each filtering type within each consumer class category.

  ○

  The THAPE uses the backtracking concept to identify other products associated with a given product that the business considers relevant.

  ○

  Its main tasks include the following:

  ✓

  Uncovering useful rules based on given requirements.

  ✓

  Exploring hidden association rules that may influence other factors.

  ✓

  Providing backtracking capabilities for a given product.

  ○

  Output: The most useful rules that best suit the business needs.

In this paper, we employ the main outcome of this framework as well as some new techniques to act as enablers for Metaverse to improve personalized shopping experience. As detailed in Section 6, from these outcomes we develop an AI framework that helps in predicting consumer behavior based on association rule learning.

3.1. Loyalty Customers Clustering

We applied the K-means clustering algorithm with k = 3 to a dataset of 29,347 active consumers from the loyalty card program repository. Figure 4 summarizes the clustering process.

Based on all the clustering results, we can summarize some properties about each class as shown in

Table 3. The clustering results summary.

	Class A	Class B	Class C
i. Entire class average spend/month	5,855,404 SR	3,224,208 SR	2,732,594 SR
ii. Number of consumers in the Class	6502	10,176	12,669
iii. Average visit/month	4.086	1.153	0.683
iv. Average spend/month (=I ÷ ii)	900.690	316.792	215.712
v. Individual average spend/visit (=iv ÷ iii)	220.4 SR	274.8 SR	315.8 SR

:

Regarding spending and visiting frequency habits, it is clear that Class A spends the most overall per month (5,855,404 SR) and visits the most frequently, averaging around four visits per month. Although Class A spends the least per visit, this suggests that they visit more often but make smaller purchases each time. This could indicate that they are more engaged with the location, due to loyalty programs, proximity, or a wider range of offerings, or that they have more disposable time and money.

Class A also has the highest percentage of Saudi nationals (62.74%), whereas other classes have less than 50%. This could be relevant if cultural or national preferences influence spending habits.

Regarding individual spending per visit, Class C spends the most per visit (315.8 SR). However, despite spending the most per visit, Class C’s overall monthly spending is the lowest. On the other hand, Class C visits the least frequently, averaging around a visit or less per month. This may indicate lower interest, limited accessibility, or different spending priorities. While they visit less often, when they do, they spend more. This might suggest they make larger, less frequent purchases. Additionally, demographic data show that Class C has the highest percentage of female individuals at 20.11%. This could potentially influence spending patterns if certain products or services appeal more to women. However, this is not conclusive evidence, as the percentage of female individuals is low. Moreover, Class C has the highest number of unemployed individuals at 21% (students or housewives). This might explain their lower overall spending and lower visit frequency, as they may have less disposable income. While unemployment likely contributes to lower spending, other factors like spending priorities and lifestyle choices may also play a role.

To sum up, Class A is the most active and highest-spending group overall, significantly impacting overall revenue, driven by frequent visits and high average spending per person. Class C spends the most per visit but visits the least, potentially due to financial constraints or different purchasing habits. The demographic data provide some clues but require further context to fully understand the spending behaviors of each class.

3.2. Interesting Association Rules

We ensured that our framework considers the interesting concept as a tunable concept, and we provided a backtracking mechanism to ensure that any interesting criteria remain applicable. As mentioned in our previous publications [17,36], we built our entire work in collaboration with the business owner of one of the largest supermarket chains in Saudi Arabia. Therefore, the interesting concept in this study was considered a case study.

First, the model was trained on a labeled dataset containing samples of association rules. In our case, based on product price and distance, the interesting rule was determined as explained in Algorithm 1 below.

Algorithm 1: interesting rule calculation

for entire class: calculate the average price of individual product    if (the price of LHS > average OR the Price of RHS > average)     Calculate distance between LHS and RHS     distance = LHS $\stackrel{\mathrm{D}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}}{\to }$ RHS     if (distance > 3)       Return interesting     else       Not interesting   else     Not interesting End

These associations involve products that are located in different sections of the store. They are interesting because of Unexpected Shopping Patterns, which represent the distance the customer needs to move from one aisle to another inside the supermarket, indicating that there is an obvious connection that goes against store organization. Store Layout Optimization suggests potential inefficiencies in the store layout if customers often buy products that are far apart in terms of layout. The store might benefit from consumer movements and provide cross-section ads. Cross-Category Demand reveals purchasing behaviors that span across different product categories (e.g., electronics and kitchenware). It suggests real-life cases that store planners might not have considered. Customer Effort and Experience imply a high level of need or complementarity if customers are willing to traverse the store to buy items together. Recognizing these patterns can help in improving customer experience by offering better placement or signage. Understanding Marketing and Promotion associations can help in designing promotions that encourage multi-category purchases, such as discounts for buying items from different sections.

For the price-based interesting association, those associations that involve high-price products are interesting because high-price products are typically purchased less frequently than low-cost items. So, this indicates Customer Behavior Insights if those products appear in strong association rules, suggesting a meaningful connection in customer purchasing behavior. Bundling and Cross-Selling Opportunities identify high-price products that are frequently bought together and can help in designing special promotions or bundle deals, increasing overall sales revenue. Premium Customer Segmentation identifies customers purchasing high-priced items that may represent a specific segment (e.g., luxury shoppers). Understanding these associations can help with personalized marketing strategies. Profit Maximization involves studying these associations, which can significantly contribute to the store’s revenue since the buying behavior of customers who purchase a cheaper product often leads to the purchase of a more expensive product. Analyzing these relationships can guide pricing strategies and inform adjustments to store layout, ultimately maximizing profit.

4. Business Point of View

During this journey with the business owner, we developed a short questionnaire to understand their perspective on our results from a business standpoint. We asked our stakeholders, including directors and managers from both the marketing department and the information technology department, to complete the questionnaire.

Table 4. Summary of business point of view.

Aspects	Concern	Remarks
Adaptability	Relevance of dimensions: responses varied regarding the relevance of these dimensions.	While some indicated that the outcome and interpretation being pursued determine their significance, others affirmed their importance in understanding consumer behavior.
Flexibility	Adapting to market trends and consumer behaviors.	Participants raised concerns about the framework’s ability to adapt to changing market trends and consumer behaviors. They emphasized the importance of flexibility in strategies to ensure continued relevance and effectiveness in analyzing consumer behavior.
Usefulness	Assessing customer purchases and evaluating rules: some participants favored this approach, emphasizing its potential to assess customer purchases and evaluate rules.	However, others suggested alternative methods, such as splitting the data into training and testing sets or applying the methodology to a sample customer and then assessing their purchases, as they believed that using a survey could introduce challenges and potential inaccuracies.
Data Integrability	Ensuring seamless integration of diverse data sources.	Some participants highlighted the need for the seamless integration of diverse data sources to enhance the accuracy and comprehensiveness of the analysis. They stressed the importance of ensuring that data from various sources can be effectively combined and analyzed to provide meaningful insights into consumer behavior.
Correctness	Achieving expected results: some participants expressed skepticism, stating that while it sounds good, the expected results may not be achieved.	A slightly different response indicated concern that changing the arrangement may confuse and frustrate customers. Another participant acknowledged the value of trying the approach but suggested applying manual filtering as needed.
Scalability	Handling increasing data volumes.	Participants expressed concerns about the framework’s scalability to handle increasing data volumes as the business grows. They emphasized the importance of ensuring that the system can efficiently process and analyze large amounts of data without compromising performance or accuracy.
Effectiveness	Enhancing virtual shopping experience: they expressed their interest and praised its effectiveness.	They highlighted that through virtual shopping, certain products can be shown when customers make a purchase, and additional product suggestions can be personalized based on their purchase history. Participants also noted that this approach may lead to increased basket size, improved purchasing habits and customer loyalty. Overall, they agreed that this approach would be relied upon, as the Metaverse makes it easier to remind users of specific items based on the association rules.

presents a conceptualization of the questionnaire in terms of seven aspects: Adaptability, Flexibility, Usefulness, Data Integrability, Correctness, Scalability, and Effectiveness. Figure 5 presents these seven aspects and how far the business cares for and attends to each aspect in a spider-web diagram. In the results section, we provide a subjective evaluation of the consumer shopping experience based on the questionnaire results, while Table 4 summarizes an objective evaluation from a business perspective.

5. Metaverse as a Platform for Retailers

The Metaverse is often described as an immersive, 3D virtual world where multiple users can interact in an online environment [3]. The Metaverse, a collective virtual shared space, is emerging as a transformative force across industries. For retailers, it presents a unique opportunity to reimagine the shopping experience [14,41]. The Metaverse offers retailers a new frontier for immersive shopping experiences. Customers can virtually see the products, explore virtual supermarkets, and interact with products in 3D, creating a more engaging and personalized retail journey [16,24,51].

The Metaverse has two main components, the virtual worlds and the avatars [14,37]. Virtual worlds are digital environments where users can interact with each other and the environment, while avatars are digital representations of users that allow them to navigate and interact within the Metaverse. Within the Metaverse, users can interact with a computer-generated environment and other users through the use of avatars.

Figure 6 presents a layered architecture for the metaverse technical stack [61,62,63,64]. (1) The distributed cloud and the decentralized cloud at the bottom of the stack provide a scalable backbone for the Metaverse enterprise. (2) The digital assets and digital cutting life cycle provide a repository for any digital asset entities and life-cycle management for the digital assets. (3-a) The blockchain ledger provides a decentralized public ledger for any digital tokens. (3-b) The IoT platform and digital twins provide a scalable platform for allowing entities to connect using various protocols and digital twins for any entity. (3-c) The immersive platform provides virtual platforms for AR/VR and representations for converting any physical representation into a virtual entity and providing enhanced AR/VR at scale. (3-d) The AI/intelligent platform provides various vertical-focus AI agents in the Metaverse, providing intelligence for various virtual entities. For example, the use cases can range from behavior analysis for virtual entities to more personalized education or buying experience for the end user. (4) Immersive apps leverage the Metaverse stack to provide various hybrid solutions; for instance, this can be the virtualized shopping experience.

Supermarket shopping in the Metaverse offers immersive experiences for consumers, with virtual aisles, interactive product displays, and virtual try-ons. Businesses benefit from increased customer engagement, data-driven insights, and the ability to create unique branding experiences. Supermarket shopping in the Metaverse offers various features and benefits from both consumer and business perspectives.

Table 5. Benefit of Metaverse from both consumer and business perspectives.

Consumer Perspectives

Business Perspectives

- Immersive Shopping Experience: users can virtually explore products in 3D, try them on using AR, and obtain a realistic shopping experience [25,65]. - Personalization: AI algorithms can analyze user data to provide personalized product recommendations [16,30]. - Social Shopping: users can shop with friends in real-time, share feedback, and make group purchases [14,15,53]. - Virtual Storefronts: businesses can create virtual stores with interactive displays, enhancing branding and customer engagement [65]. - Digital Ownership: blockchain technology enables secure ownership of virtual assets, such as limited-edition digital goods [3,27].

- Enhanced Customer Engagement: businesses can engage with customers in new and interactive ways, leading to increased brand loyalty [38,51]. - Cost-Effective Marketing: virtual events and promotions within the Metaverse can reach a global audience at a fraction of the cost of physical events [19,49]. - New Revenue Streams: businesses can sell virtual goods and offer virtual services within the Metaverse [37,50]. - Data Analytics: the Metaverse provides valuable data on user behavior and preferences, helping businesses make informed decisions [17,20]. - Global Accessibility: businesses can reach customers worldwide without physical limitations, expanding their market reach [10,18].

summarizes these aspects for a comprehensive overview.

While the Metaverse is still in its early stages, its potential to revolutionize retail and other industries is undeniable [66,67]. By embracing these technologies and innovative approaches, businesses can unlock new opportunities and create truly immersive experiences for their customers [68].

6. Proposed Architecture: THAPE as an Enabler for Personalized Shopping in Metaverse

This paper aims to empower retail in the Metaverse through leveraging consumer behavior analysis and contributing to the field of personalized shopping experiences. In this paper, we propose an architecture that seamlessly integrates the physical and digital worlds, such that consumer behavior in the real world is analyzed to feed the Metaverse.

Figure 7 illustrates the feeding flow from the real-world supermarket environment to the virtual realm in the Metaverse, showcasing the pivotal role that our engine plays in elevating the shopping experience through customization. The THAPE works as follows:

• From the real-life environment, the THAPE takes the following two data sources as a base to start:

  ○

  Personal information from the Loyalty Program Repository;

  ○

  Sales transactions of individuals.

• The THAPE, for the Metaverse, carries out the following:

  ○

  Starts Big data processing including data transformation, feature selection, etc.

  ○

  Applies Association Rule Learning to find our product’s relations and connections.

  ○

  Re-calls Clustering to identify the appropriate cluster to which the consumer belongs.

  ○

  Applies Classifications to classify association rules into “interesting” and “not interesting”. Then, the THAPE implements a tunable filtering and sorting process.

• The THAPE feeds the Metaverse with real-time predictions and recommendations that empower virtual supermarkets through the following:

  ○

  Shelf Re-arrangement: product rearrangement based on user selection.

  ○

  Individual Assessment: users start receiving assessments based on their behavior.

  ○

  Personalized Shopping: users interact but each has their own view.

Data Labeling and Classification

Data labeling is a form of data processing that includes price and distance labeling:

Price labeling refers to the unification process, as a product’s price in sales transactions is scattered across different quantities and sizes. Therefore, we standardized the dataset by identifying a unique set of items with a corresponding price calculated as the net price to ensure consistency in price labeling. Next, we exported the D4 class name of each item, which includes the product name without any details of product size or weight. Moreover, we addressed products with a quantity of less than one, which typically represent weighted items such as fruits, vegetables, cheese, olives, pickles, and so on. To ensure consistency and avoid treating them as separate products, we standardized prices based on the price per kilo. Finally, all prices were rounded to two decimal places to maintain consistency and precision in the dataset.

Distance labeling refers to the value of the physical distance between the locations of products involved. This concept is based on cross-sections, where the distance value between products increases as the products come from different sections. Therefore, we calculated the distance as the number of sections between any two given products.

For the classification, we applied three classifiers to find the best fit for our data. These classifiers were random forest, and naïve Bayes. We compared their performance metrics and selected the best one to classify our data. The classifier model has been evaluated using standard performance metrics such as accuracy, confusion matrix, and classification error.

The random forest classifier achieved the highest overall accuracy, so we built our classifier model as shown in Figure 8 below.

This classification process categorizes the association rules into “interesting” and “not interesting” for each consumer cluster. Subsequently, a tunable filtering and sorting process is implemented to rank these rules from most to least interesting based on attributes such as price, confidence level, and the distance between the involved products.

The THAPE delineates the diverse stages of the virtual shopping journey, starting from product discovery, with a strong emphasis on user interaction and personal recommendations crafted to cater to individual preferences. It concludes with buying behavior recorded iteratively, which feeds back to improve the overall experience.

This allows users to explore virtual stores that mirror their behaviors in real-world layouts and product offerings. AI-powered personalization enables tailored product recommendations and virtual try-on experiences. This integrated approach enhances customer experience, improves operational efficiency, and unlocks new revenue streams for retailers.

7. Consumer Behavior Prediction

7.1. Experiment Design

We developed a pilot study centered on a survey methodology. This survey was constructed utilizing data accumulated over a span of three months, comprising sales transactions and consumer profiles sourced from a specific branch of supermarkets in collaboration with the business owner. To ensure a diverse respondent pool, the survey forms were distributed among a randomly selected group of consumers who frequently visited the same branch where transaction data were collected for the predictive model. This approach resulted in approximately 1000 participants, with 800 qualified responses used for the pilot analysis. The key criterion focused on various factors, with particular emphasis placed on the frequency of visits to the supermarket within a given month. This meticulous approach allowed us to evaluate our results and calculate the prediction accuracy as well as gather valuable insights into consumer behavior patterns and preferences within the context of retail shopping, paving the way for a deeper understanding of consumer interactions with the supermarket environment.

The experiment was divided into three distinct phases as shown in Figure 9. In the initial phase, the THAPE extracted sales transaction data from the repository of historical data to generate predicted lists for each customer class—A, B, and C. Following this, the second phase involved market research, where a survey was conducted to validate the accuracy of the predicted lists. A market research survey was designed based on these lists to discern consumer predictions and preferences. The participating consumers in this survey were randomly selected from the same branch. This phase ultimately led to the derivation of validation results and the creation of actual lists in the final order. In the final phase, an accuracy calculation was performed by comparing the actual lists with the predicted lists. This comparison allowed for the calculation of accuracy from both a horizontal and vertical perspective, providing insights into the effectiveness and reliability of the prediction model.

7.2. Question Design

Our THAPE extracted sales transaction data from the Repository of Historical Data to generate predicted lists of association rules for each customer class—A, B, and C—based on two criteria: distance-based and price-based association rules. In this section, we explain how we designed the survey questions using the THAPE-predicted list for Class A (as an example). Below is the list of the top 10 distance-based association rules for Class A. The rules are ordered from most to least interesting (ranked 1 to 10). The first rule has the highest confidence value, and the two associated items have the maximum distance, followed by the rest in descending order.

1.

LHS₁: Dishwashing Liquid Fairy  →  RHS_1-1: Cucumber vegetables in kg

2.

LHS₂: Tea Bags Lipton        →  RHS_2-1: Water stainless steel container

3.

LHS₃: Tomato Paste In Tetra PC   →  RHS_3-1: Top Loading- Tide

4.

LHS₄: Corn Oil Mazola       →  RHS_4-1: Top Loading Ariel

5.

LHS₅: Dishwashing Liquid Fairy  →  RHS_5-1: Pure animal ghee

6.

LHS₆: Chicken Breast Sadia     →  RHS_6-1: Fine baby diapers

7.

LHS₇: Chicken Stock Maggi     →  RHS_7-1: Plain Cream- Puck

8.

LHS₈: Vegetable Oil Al Arabi     →  RHS_8-1: Sun protector cream

9.

LHS₉: Fresh Potato        →  RHS_9-1: Eggs

10.

LHS₁₀: Fine Sugar in Bag       →  RHS_10-1: Coca Cola Pepsi

Similarly, below is the list of the top 10 price-based association rules for Class A. These rules are also ranked from most to least interesting. The first rule has the highest confidence value and the largest price difference, followed by the rest in descending order.

11.

LHS₁: Fabric Softener Comfort   →  RHS_1-1: Frozen Chicken 10 × 1200 g

12.

LHS₂: Liquid Tahina Al Jameel      →  RHS_2-1: Coffee Beans Harari

13.

LHS₃: Facial Tissue           →  RHS_3-1: Steam Iron

14.

LHS₄: Baladi Veal Baladi Veal     →  RHS_4-1: Electric water heater

15.

LHS₅: Tea Biscuits          →  RHS_5-1: Top Loading Detergent

16.

LHS₆: Regular Shampoos      →  RHS_6-1: Baby Diaper Mega Pack

17.

LHS₇: Top Loading Detergent    →  RHS_7-1: Hand Blender

18.

LHS₈: Raw Flour White Flour    →  RHS_8-1: Chicken Fillet Al Kabeer

19.

LHS₉: Mozzarella Shredded    →  RHS_9-1: Orange Tang

20.

LHS₁₀: Electric Juicer       →  RHS_10-1: Electric Deep Fryer

Then, we began constructing the survey questions by selecting three different RHSs from the THAPE results for each LHS in the list. For the distance-based questions, we ensured that the three RHSs had varying distances relative to the LHS. The optimal option was the one that included the farthest RHS from the given LHS. To clarify, let us take the first association rule from the distance-based list as an example. The rule is as follows: LHS₁: Dishwashing Liquid Fairy→RHS_1-1: Cucumber vegetables in kg. Figure 10 illustrates the three selected RHSs along with their respective distance values, measured by the number of sections between them.

Similarly, for the price-based questions, we ensured that the three RHSs had varying price differences. The optimal option was the one that included the most expensive RHS compared to the other RHSs. To clarify, let us take the first association rule from the price-based list as an example. The rule is as follows: LHS₁: Fabric Softener Comfort→RHS_1-1: Frozen Chicken 10 × 1200 g. Figure 11 illustrates the three selected RHSs for the given LHS:

7.3. Vertical and Horizontal Accuracies Meanings

In this study, we tested how accurate the THAPE is in predicting association rules using two different types of accuracy as follows:

Vertical Accuracy: This refers to the accuracy of ordering the ten association rules in each list from the filtering and sorting stage of the framework, compared to the resulting order from the survey. As shown in Figure 12, we compared the predicted order of the interesting rules with the actual order identified from consumer survey responses.

Horizontal Accuracy: This refers to the extent to which the optimal RHS_1-1 is chosen among RHS_1-2 and RHS_1-3 presented in the question. Horizontal accuracy is calculated by dividing the number of respondents who selected that relationship e.x LHS₁→RHS_1-1 by the total number of participants who answered the question. This metric represents the accuracy of our prediction, which will be compared with the accuracy of the classifier.

After calculating the prediction accuracy for both price-based and distance-based questions in horizontal and vertical view, we discuss and present the expected revenue per each consumer’s class (in terms of monetary value) if customers follow the prediction based on the results of the experiments. We then estimate the increased sale value for some selected products by tracking the number of transactions where the prediction is proven.

8. Result and Analysis

In this section, we calculate and present the horizontal and vertical accuracy for Class A, Class B, and Class C for both distance-based and price-based questions of the survey.

Table 6. Vertical and horizontal accuracy summary of price-based questions.

Relations (Prediction)	Class A		Class B		Class C
	Horizontal Accuracy	Vertical Accuracy	Horizontal Accuracy	Vertical Accuracy	Horizontal Accuracy	Vertical Accuracy
LHS1→RHS1-1	43.3%	80%	43.2%	60%	30%	90%
LHS2→RHS2-1	28.5%	30%	59.3%	90%	20%	20%
LHS3→RHS3-1	40.2%	80%	32.2%	60%	50%	80%
LHS4→RHS4-1	32.1%	60%	28.8%	60%	50%	70%
LHS5→RHS5-1	43%	70%	28%	50%	40%	80%
LHS6→RHS6-1	24.4%	60%	54.2%	70%	30%	70%
LHS7→RHS7-1	51.3%	70%	44.1%	70%	50%	60%
LHS8→RHS8-1	51.3%	70%	41.5%	80%	40%	80%
LHS9→RHS9-1	47.2%	80%	58.5%	30%	50%	50%
LHS10→RHS10-1	47.2%	60%	23.7%	90%	70%	80%
Average	40.8%	66%	41.4%	66%	43%	68%

below summarizes the accuracy of the price-based association rules.

Table 7. Horizontal and vertical accuracy summary of distance-based questions.

Relations (Prediction)	Class A		Class B		Class C
	Horizontal Accuracy	Vertical Accuracy	Horizontal Accuracy	Vertical Accuracy	Horizontal Accuracy	Vertical Accuracy
LHS1→RHS1-1	24.9%	20%	30.5%	50%	51.3%	100%
LHS2→RHS2-1	43.5%	90%	30.5%	50%	47.54%	90%
LHS3→RHS3-1	41.5%	90%	49.2%	100%	33.2%	40%
LHS4→RHS4-1	48.7%	80%	57.6%	70%	36.6%	60%
LHS5→RHS5-1	33.4%	80%	20.3%	60%	44.15%	90%
LHS6→RHS6-1	36.8%	90%	55.9%	60%	47.16%	80%
LHS7→RHS7-1	37%	70%	44.9%	70%	47.92%	50%
LHS8→RHS8-1	27.5%	80%	34.7%	70%	39.24%	90%
LHS9→RHS9-1	16.1%	90%	22.9%	90%	46.03%	60%
LHS10→RHS10-1	50.8%	10%	20.3%	100%	31.69%	100%
Average	36%	70%	36.7%	72%	42.9%	76%

below shows the horizontal and vertical accuracy of the distance-based questions for all classes:

This study demonstrates an acceptable level of accuracy in both the ordering of association rules from most to least interesting (vertical view) and the prediction of products within each rule (horizontal view). The alignment of consumers’ buying interest in both price-based and distance-based questions is strongly supported by the data predictions. However, stronger alignment was observed with distance-based predictions.

Horizontal accuracy measures the accuracy of predictions based on the initial rule classification using a random forest classifier. The comparison between horizontal accuracy and classifier accuracy shows good agreement. Class C exhibits higher accuracy in both price-based and distance-based questions, as well as in the classifier results, indicating that Class C has the highest accuracy according to the classifier.

Vertical accuracy measures the accuracy of ordering association rules from most to least interesting, which is more important than horizontal accuracy. Vertical accuracy is more important as it highlights a significant aspect of the rule, which is the left-hand side (LHS). This guides consumer behavior and indicates which products to use in bundling. The right-hand side (RHS) is a consequence and is not as critical as the LHS. Vertical accuracy is responsible for selecting the most interesting rules, focusing on the LHS.

The average vertical accuracy ranges from 76% to 66%. This high accuracy, in comparison to horizontal accuracy, supports our framework and THAPE design. The tunable filtering and sorting act is an advanced stage following the classifier to enhance the interpretability of association rules. This further validates the effectiveness of our add-on stage for ordering association rules, enabling business owners to identify the best products to initiate their actions.

High vertical accuracy proves that using confidence levels was a good approach.

The average horizontal accuracy ranges from 43% to 36%, measuring the support of RHS_{X-1 where x is from 1 to 10 in each class.} Each given LHS has three RHSs, achieving an average accuracy across all classes exceeding one-third, approximately 33.3%, assuming an equal fair distribution probability. This is particularly noteworthy since all RHSs are selected from the framework results, indicating a high likelihood of selection as they are predicted based on consumer behavior.

Our prediction of interesting orders (vertical accuracy) is more precise when using distance-based relations rather than price-based relations. Our model performs better with distance-based relationships, indicating our ability to detect more subtle patterns. The bundling of LHS and RHS becomes more challenging when incorporating distance-based products compared to price-based bundling. This is crucial because relationships based on distance tend to be more nuanced, especially when analyzing purchasing transactions. Our prediction is more accurate in identifying hidden distance-based relationships, which are both more significant and less obvious.

Class C emerges as the most predictable class, exhibiting the highest vertical and horizontal accuracy:

• This may be attributed to its large consumer base. As discussed in the previous analysis section, the clustering output indicates that Class C captures the largest number of consumers.
• The behavior of consumers in Class C—who tend to spend more but visit less frequently—is more predictable than that of consumers in Classes A and B. This is advantageous from a business perspective, as it signals strong potential for targeted marketing strategies.
• Class C is particularly predictable, especially in price-based predictions. This can be explained by the professionalism of Class C consumers, who exhibit the highest average spending per visit per individual.
• Visiting less frequently may indicate a more concentrated, strategic, and planned approach to grocery shopping, leading to predictable behavior in the basket analysis.

Buying a cheap product is positively correlated with purchasing expensive products, as we noted in terms of price-based vertical accuracy.

Also, based on the distance-based horizontal accuracy, we could say that buying a product from one section leads to cross-sectional buying behavior which can be utilized in different marketing strategies.

All results and findings can be utilized in marketing strategies and product arrangements.

9. Discussion and Future Work

In summary, the Metaverse represents a significant evolution in interactive experiences, with potential applications in retail. It offers immersive and personalized shopping experiences that enhance consumer engagement and drive sales. Retailers who effectively harness the potential of Metaverse may gain a competitive advantage in the rapidly changing landscape of consumer behavior. By utilizing the capabilities of the Metaverse, consumers engage with products and make purchasing decisions in a groundbreaking way.

Our proposed architecture: the THAPE as an enabler for personalized shopping in the Metaverse that contributes to enhancing the personalized shopping experience within virtual supermarkets, serving as a key enabler in the Metaverse environment. Acting as a crucial feeding engine, the THAPE bridges the gap between the physical world and the Metaverse by providing real-time predictions and valuable insights. This innovative engine integrates various AI techniques, including big data analysis, association rule learning, clustering, and classification, to drive its predictive capabilities and optimize the customization of user experiences.

The Metaverse, with our proposed architecture, allows individuals to interact with products and connect with other consumers in a virtual space, fostering a sense of individuality in their shopping experiences. Users can explore products, share insights, and make informed decisions in a dynamic and socially engaging digital realm. The advent of Metaverse technology has the potential to reshape the retail landscape, offering a transformative shopping experience that bridges the gap between physical and digital realms.

Additionally, the revelation of cross-sectional buying behavior in response to purchases from specific sections highlights the practical value of these findings in optimizing product arrangements. The comprehensive validation of these questions not only enhances our understanding of consumer behavior but also offers actionable insights for businesses aiming to tailor their strategies effectively.

By implementing our predictive architecture in supermarket sales within the Metaverse, we anticipate a transformative revenue impact on consumer behavior and purchasing patterns. For instance, consider the most predictable Class C, where Tide is currently associated with tomato paste, resulting in sales of SAR 41,860 based on existing data. With the application of our prediction model, we project a significant surge in sales, with Tide sales expected to soar to SAR 364,180. This extrapolation exemplifies the potential of our predictive analytics in optimizing product placements and rearrangements in the Metaverse, enhancing cross-selling opportunities and ultimately driving substantial revenue growth within the virtual supermarket environment.

For future work, we recommend several avenues for further exploration: Enhancing Metaverse Enablers, which is crucial to enhance enablers within the Metaverse to broaden the scope of studies and uncover more subtle relationships. This could involve integrating advanced analytical tools and methodologies that facilitate a deeper understanding of user interactions. Through Optimization in Rule Generation, we recommend incorporating more optimization stages in the generation of association rules. This enhancement could lead to more efficient and effective rule mining, ultimately improving the accuracy of insights derived from user data. Integrating Multiple Predictive Aspects: It would be beneficial to build a framework that encompasses multiple predictive aspects and applies optimization techniques among them. This approach could yield a more comprehensive understanding of user behavior and preferences. Beyond Personalized User Experiences: Association rules analyze patterns in user transactions to identify relationships between items. We recommend developing sophisticated algorithms that not only suggest relevant products but also adapt over time as user preferences evolve. This would create a more tailored shopping experience, akin to a recommendation system. Exploring Alias Sales in the Metaverse: We suggest studying the relationship between alias sales transactions in Metaverse supermarkets and the interactions within aisles. Understanding how these aliases impact consumer behavior could provide valuable insights into shopping patterns and preferences. Sequential Traceability of Behavior: We propose investigating associations as a means of sequentially tracing individual behavior. By tracking user movements within the metaverse, we can exploit these sequences to develop diverse strategies tailored to user engagement and experience.

By pursuing these avenues, we can further enrich our understanding of consumer behavior in the Metaverse and improve the overall shopping experience.

10. Conclusions

In this paper, we propose a Metaverse-driven shopping architecture designed to empower the retail sector through comprehensive consumer behavior analysis, facilitating personalized shopping experiences. Our contributions to Metaverse supermarket shopping are as follows: Presenting a comprehensive review of the Metaverse in the retail sector and its key enablers. Proposing the THAPE as a framework enabler for personalized shopping in a Metaverse environment. Evaluating the accuracy and benefits of the THAPE based on a pilot study conducted in collaboration with a major retailer in Saudi Arabia.

The validation analysis of our framework, based on real-world consumer responses, yielded promising accuracy rates across various association rules. Notably, it highlighted significant vertical and horizontal accuracy in Class C, making it the most predictable consumer class. The underlying reasons for this are discussed in detail in the Results and Analysis Section. Moreover, vertical accuracy outperformed horizontal accuracy. In association rule learning, the left-hand side (LHS) of a rule is particularly important. In our case, vertical accuracy was more significant, as it determined the LHS of the rule. Regarding distance- and price-based associations, the accuracy results indicated better performance in distance-based association predictions, which is especially relevant for product bundling.

Our in-depth analysis provides insights that enable enterprises to tailor Metaverse shopping environments according to consumer preferences. This approach not only enhances customer satisfaction but also increases conversion rates by leveraging accurate product predictions and customization features. Furthermore, our findings underscore the potential of Metaverse technologies in retail, emphasizing the importance of understanding consumer behavior in virtual settings. We recommend further research into optimizing the discovery process of association rules to better align business interests with consumer preferences. By adopting this methodology, retailers can enhance customer engagement, differentiate themselves in a competitive market, and foster greater customer loyalty.