Investigating continuous intention to use metaverse in higher education institutions: a dual-staged structural equation modeling-artificial neural network approach

The context of higher education is constantly shifting as a result of the proliferation of new technologies and the spread of globalization. The majority of educational institutions, regardless of their size or shape, battle to compete by enhancing their technological skills (Khalid et al., 2018). Recently, there have been many changes to the higher education sector scene within and after the pandemic. Many HEIs globally, including Egypt, have transitioned to virtual learning. The challenge that COVID-19 poses to the world's education sector is profound (He et al., 2022). Many education systems have recently been forced to take the learning process and other operations remotely (Comelli et al., 2021). This path results in innovative strategies to cross digital gaps and overwhelm other issues (Salloum et al., 2023).

Metaverse refers to the online virtual world (Metwally et al., 2024). According to Chua and Yu (2024), this new frontier has the potential to be a game-changer for businesses all over the world. It is a magical zone where real life and virtual reality mix to produce an experience that is immersive and entertaining for numerous users (Chen et al., 2023a, 2023b; Mystakidis, 2022). Researchers have previously associated the term "metaverse" with technologies such as augmented reality (AR), extended reality (XR), virtual reality (VR), and mixed reality (MR). These technologies make use of a three-dimensional (3D) virtual world to provide users with an experience that is both immersive and collaborative (Chen et al., 2023a, 2023b; Tan et al., 2023). During the years preceding the COVID-19 pandemic, the foundation of the pandemic was established in the virtual world as a means of overcoming constraints brought about by individual demands and responsibilities involved in extracurricular activities. Additionally, the influence of online education may have been of a higher quality; nevertheless, it was hampered by a number of issues, such as inadequate technological conditions, infrastructure, software applications, and other issues (Lu et al., 2024; Park & Kim, 2022).

Many issues that need to be addressed immediately in online teaching have surfaced. Lack of context, engagement, and participation in online instruction must be addressed quickly (Kaddoura & Al Husseiny, 2023; Zhang et al., 2022). Recent advances in metaverse technology have greatly benefited online education (Wang & Shin, 2022). Because Generation Z (born after 1995) is so different from their predecessors, they are particularly open to the (Park & Kim, 2022). Metaverse can be utilized to teach in numerous circumstances (Kim et al., 2022). Many firms, including Meta, Roblox, Zepeto, and others, are developing metaverse classroom tools. Curriculum creation includes the metaverse, making it more engaging for pupils. While metaverse technology has immense potential in the classroom, its adoption is still young. Free mobile platforms are worse for education (Zhang et al., 2022). Thus, users might be drawn to the metaverse's education ecosystem by studying how the application's platform affects their objectives. It may also fill the requirement for certain instructional technology (Xu & Impagliazzo, 2024).

As a result of the metaverse system's relative novelty, a limited amount of research has been conducted on the acceptability and deployment of the system by users in industrialized countries. According to Tlili et al. (2022b), Metaverse looks to be beneficial to its users in terms of facilitating better classroom administration. As a consequence, there is a significant possibility that this study will shed light on the elements that influence the adoption of metaverse technology by information technology. In this study, we hope to contribute to the existing body of literature on educational technology application platforms in formal educational environments in three different ways. This will allow us to fill the gaps that have been identified. Initially, previous research has focused on the intention to implement various platforms for educational technology platforms. The majority of research focuses on e-learning (Nayak et al., 2022), mobile learning (Alowayr, 2022), educational mobile games (Tlili et al., 2024), learning management (Al-Nuaimi & Al-Emran, 2021; Tlili et al., 2022a), augmented reality and virtual reality technologies (Jang et al., 2021; Papakostas et al., 2022), and social media services (Al-Rahmi et al., 2021; Yu, 2020). However, Metaverse educational technology differs from traditional educational technology in many ways, including pedagogical aims, intended learners, necessary technological infrastructure, course content, and accessibility to technical (Park & Kang, 2021; Tlili et al., 2023). Therefore, further research into the factors that encourage users to participate in the metaverse is consistently required.

Second, artificial neural networks (ANN) are used in conjunction with partial least squares structural equation modeling (PLS-SEM) to rank the normalized importance of the predictors in order to validate the results of PLS-SEM (Lee et al., 2023; Tan et al., 2014). As a result, this research not only contributes significantly to our theoretical understanding of the uses of virtual learning in educational contexts, but it also offers practitioners vital insights. Thirdly, this research makes a contribution to the existing body of knowledge by combining the three theories of self-determination theory (SDT) and the technology acceptance model (TAM). It provides a comprehensive explanation of the elements that influence the continual intention of Egyptian students to utilize the metaverse. A lack of rigorous theoretical knowledge and empirical research studying the elements of ongoing intention to employ metaverse in learning is the specific problem that this study addresses. In conclusion, this study addresses the specific problem discussed in the previous sentence. As a result, the primary research question that is motivating this study is as follows: What are the key factors and the most crucial ones that influence students' continuing intention to use the metaverse in learning?

The metaverse's classroom benefits have piqued academic curiosity. The metaverse definition of education is the main subject of certain studies. Metaverse technology creates a novel learning environment (Morganti & Bartolomei, 2024; Wang & Shin, 2022). The distinct advantages of the educational Metaverse have aroused the curiosity of scholars (Wang et al., 2024). Initially, there is a body of work that primarily addresses the educational context when discussing the Metaverse. Metaverse technology has created a new kind of learning environment that combines four different kinds of augmented reality, virtual reality, life logging, and a mirror world (Salloum et al., 2023). Additionally, other studies shed light on the features of the Metaverse as they pertain to educational settings. The 5Cs—currency, continuity, canon, creator, and connectivity—were put forth by Go et al. (2021) as characteristics of the Metaverse. Compared to traditional classroom instruction, screen-based remote instruction, and instruction in the Metaverse, the former three have many limitations in terms of time and space, while the latter two have many advantages, such as strong interaction, virtual identity, immersive experience, open and accessible creation, and thorough evaluation of teaching (Zhang et al., 2022).

Thirdly, various studies have addressed the universe's most important educational resources and technologies (Büyüközkan & Mukul, 2024). Tlili et al. (2022b) classified a wide range of technologies and approaches, including immersive modeling and simulation, game applications, artificial intelligence (AI), education, mobile, sensors, and wearable devices. In addition, cloud computing, artistic intelligence, blockchain, 5G, graphics processing units (GPUs), VR, AR, photography 3D engines, and extended reality (XR) were highlighted by Contreras et al. (2022) as having a supporting role in the enhancement of the educational Metaverse. Fourth, a different angle describes the benefits and ways in which Metaverse alters the educational setting. In their seven-pronged analysis of the Metaverse's beneficial effects on education, Lin et al. (2022) highlight its capacity to improve content visualization, boost learning efficiency, release previously limited educational resources, and reduce educational costs and risks. In addition, pupils now have more freedom and opportunity to think creatively because of the educational application of Metaverse software (Kye et al., 2021).

Fifthly, there was an examination of the scholastic situation and an evaluation of the universe's impact. Some of the many potential case-based educational applications described by Hu et al. (2022) include an open university, an immersive curriculum, a learning tool, a twin campus, and a virtual school, among others. Wang and Shin (2022) emphasize the benefits of the Metaverse in terms of simplicity, interactivity, immersion, usability, and interest in educational applications through user analysis of the usage effect of the spatial system learning platform. Investigating how virtual teaming affects education was the focus of Jovanović and Milosavljević (2022). Finally, the Metaverse has raised concerns among certain academics who worry it would disrupt traditional classroom practices. Wang and Shin (2022) argue that there are ethical and privacy concerns with implementing Metaverse technology into the classroom. Problems with implementing the technology, insufficient computer resources, and potential risks to students' physical and mental health are also factors to be taken into account (Hu et al., 2022). Finally, the metaverse has some academics worried that it will ruin classroom instruction. Wang and Shin (2022) argue that there are ethical and privacy concerns with incorporating metaverse technology into the classroom. When it comes to utilizing the metaverse for educational purposes, Kaddoura and Al Husseiny (2023) highlighted important obstacles, ethical concerns, and possible dangers. Problems with implementing the technology, insufficient computer resources, and potential risks to students' physical and mental health are also factors to be taken into account (Nguyen et al., 2024). Nevertheless, there is a dearth of scholarship that investigates the factors that influence students' long-term intentions to use Metaverse for learning. Therefore, this research sets out to address such a knowledge gap.

Furthermore, prior research on educational technology has made use of the TAM or a combination of the TAM with numerous theories (such as ECM, TPB, IDT, TTF, or UTAUT) (Al-Rahmi et al., 2021; Guo et al., 2022; Huang et al., 2020). Researchers have found that TAM and SDT models align well with data on the uptake of technology for online education (Racero et al., 2020). Furthermore, the PLS-SEM is used by the vast majority of empirical studies investigating the goals of educational technology. However, a few studies have discovered the integration of PLS-SEM and ANN. ANN is typically used to handle nonlinearity and multicollinearity well and provide highly accurate predictions (Abdullah et al., 2022). Finally, according to the literature, the integration of TAM with SDT to examine metaverse in learning is unknown. This study contributes by predicting students' continuing intention to use metaverse through TAM and SDT.

TAM and SDT are both incorporated into the model that is being developed by the current research. The continuing intention to use the metaverse platform is impacted by both PU and PEOU, according to the findings of TAM. Within the framework of SDT, it has been presented as a proposition that autonomy, relatedness, and competence have a major impact on PU and ease of use. Furthermore, the research model that has been proposed implies that the factor of autonomy plays a role in the continuing intention to use the metaverse platform. The research model that has been proposed is shown in Fig. 1.

Full size image

The application of a multi-analytical technique was accomplished by combining multiple PLS-SEM and ANN. Due to the nonlinear interactions that exist between the independent factors and the outcome variables, we conducted an additional analysis with multilayer perceptron (MLP) and artificial neural networks (ANN) in order to rank the normalized importance of the predictors. This was done in order to confirm the findings of the PLS-SEM. In light of the fact that the PLS-SEM can only be used to investigate the validity of linear hypotheses and cannot take into consideration the nonlinearity of correlations, the utilization of a technique that operates in two stages would be complimentary. The artificial neural network (ANN) is unable to test hypotheses due to its "black box" function, despite the fact that it is capable of recognizing nonlinear correlations (Hew & Syed Abdul Kadir, 2016; Karkonasasi et al., 2018). Additionally, the ANN is resistant to noise, outliers, and limited sample sizes, among other factors. It is also capable of supporting non-compensatory models, which are models in which a decrease in one element does not always necessitate an increase in another factor (Lee et al., 2023). For the purpose of carrying out the ANN analysis, we utilized IBM SPSS 26. In the succeeding phase, we used the important factors of the latent construct scores that were produced using PLS-SEM as input neurons for the ANN model. This was done in the same manner as Leong et al. (2020) and Abdullah et al. (2022). A feed-forward–backward-propagation (FFBP) approach, in which inputs are given in a forward channel, and the estimated errors flow in a backward direction, can be utilized to train the system to predict the outcomes of the study (Liébana-Cabanillas et al., 2017). This algorithm allows for the system to be taught to predict findings from the study. A multilayer perceptron with hyperbolic tangent activation functions was selected for the hidden layer and the identity function output. This decision was made in light of the fact that the output is continually scaled. On the other hand, we only employ the sigmoid activation function in situations when the output layer is reliant on a binary or multicast categorical function (Taufiq-Hail et al., 2021) (Figs. 3, 4, 5).

Full size image

Full size image

Full size image

By following the recommendations of Leong et al. (2020), we used 90% of the samples for the training phase, and the remaining 10% were used for the testing procedure. For the purpose of avoiding the possibility of overfitting, we employed a ten-fold cross-validating strategy and computed the root mean square of errors (RMSE) (Wong et al., 2020). According to Taufiq-Hail et al. (2021), several learning rounds have the potential to enhance the accuracy of predictions while simultaneously lowering the number of errors that occur.

Table 7 demonstrates that the regression mean square error (RMSE) values of the training and testing techniques are relatively low. The values for models A, B, and C are 0.552, 0.612, and 0.477, respectively. That being said, we are able to demonstrate that the model fits the data extraordinarily well while maintaining a high level of forecast accuracy. Our findings indicate that the ANN model is capable of accurately predicting PU with a 62.7% accuracy, predicts PEOU with an accuracy of 24.4%, and predicts CI with an accuracy of 58.8%. We calculated the R2 of the ANN model using a methodology that is comparable to the one that was utilized by Hew and Syed Abdul Kadir (2016) and Leong et al. (2020). After the H (1:2) hidden neuron, which is an inhibited cryptic neuron, the H (1:2) hidden neuron is the most significant for PU as the output layer, as shown in Table 8 (Model A). As can be shown in Table 9 (Model B), the H (1:1) hidden neuron is the most significant hidden neuron for PEOU in the output layer. This is followed by the H (1:2) hidden neuron, which is an indirect cryptic neuron. The hidden neuron of H (1:1) is the one that contributes the most to CI as the output layer, as seen in Table 10 (Model C). This is followed by H (1:2), which is an inhibited cryptic neuron. H (1:3), on the other hand, is an example of a cryptic neuron that has been repressed by the process of suppression. Considering that every single non-zero synaptic weight is connected to at least one hidden neuron, it can be concluded that every single ANN model utilized in this inquiry possesses an adequate amount of predictive significance.

By dividing the relative value of these neurons by the highest importance and expressing the result as a percentage, we were able to compute the normalized relevance of these neurons through the use of sensitivity analysis (see Table 11) (Abdullah et al., 2022). According to the findings, the most significant predictors for PU (Model A) are relatedness and PEOU (100%), followed by authority (69%) and competence (48%). Similar to the previous example, the most significant predictor of PEOU (Model B) is the degree of autonomy (100%). One hundred percent of the time, PU is the most significant predictor for CI (Model C).

When it came time to compare the outcomes of PLS-SEM and ANN, the relative rankings of each method were done. The normalized relative importance in ANN was compared to the path coefficients in PLS-SEM, which permitted this to be accomplished. Both of the evaluations are in line with the three ANN models (A, B, and C) that are presented in Table 12.

This study's primary objective was to investigate the factors that influence the intention of Egyptian university students to use the metaverse as a learning platform on a continual basis. An integration of two theoretical models, namely TAM and SDT, has been successful in achieving this research. The results that were acquired from the empirical survey are the ones that are presented in the discussion section. We discovered evidence that supports the seven additional hypotheses that were provided for the model. There is a favorable relationship between autonomy and relatedness in both PU and PEOU. As a result, autonomy and relatedness are major determinants of the outcome. Cortez et al. (2024), Wang and Shin (2022), and Racero et al. (2020) came to the same conclusion, which is consistent with our outcome. In light of this discovery, it can be deduced that the metaverse, when utilized as a learning platform, becomes advantageous and simple to employ if it possesses the qualities of relatedness and autonomy. On the other hand, competency does not have any influence on PU and PEOU. Therefore, expertise is not a significant factor in predicting this. As a result of the fact that Jeno et al. (2017) and Liaw and Huang (2015) discovered that competence was a significant component in PU and PEOU, this finding is in direct opposition to their investigations. Possible explanations for this disparity include the fact that Egypt and other countries have different cultural norms and traditions. Teachers, parents, or private tutors are the primary sources of support for kids in Egypt, for instance. As a consequence of this, they might not feel comfortable utilizing the metaverse as a platform for information acquisition. Likewise, autonomy and PU are seen as strong predictors. If students have the sense that they are valuable and independent, they are more likely to continue using the metaverse as a learning tool. These results are in line with the findings that were produced by Al-Adwan and Al-Debei (2024), Kye et al. (2021), and Racero et al. (2020). The unanticipated finding of this study, on the other hand, demonstrates that PEOU was not a major predictor in their ability to influence the continuing intention to use the metaverse as a learning platform. Wiangkham and Vongvit (2024) came to the same conclusion, which is consistent with this outcome. This is the first time that the majority of pupils have been required to utilize the metaverse; therefore, they do not have any prior understanding or experience with it.

Further, the dual analytic strategy, which consists of PLS-SEM and ANN, was utilized in order to improve and offer new insights into the findings. A comparison of the strengths of both methodologies was made in order to evaluate the linear (PLS-SEM) and nonlinear (ANN) interactions. The findings were further magnified to address the study questions by the causal links that were verified by PLS-SEM and the predictive power of the ANN. A two-staged strategy of PLS-SEM with a nonlinear, no compensatory neural network model was utilized in this research project, which resulted in the effective verification of the significant factors of CI in the utilization of metaverse in learning. One hundred percent of the normalized importance is assigned to the variables of relatedness, PEOU, autonomy, and PU. These are the most important predictors. CI's use of the metaverse may be predicted with a level of accuracy of 58.8% by the ANN model. Not only that, but the ranking comparison between the findings that were given by ANN and those that were provided by PLS-SEM reveals that the three ANN models (A, B, and C) are consistent throughout both assessments (see Table 12).

Through the conceptualization and empirical testing of a model that blends TAM and SDT theories, the current research contributes to the expansion of the existing body of knowledge. University students were given a questionnaire to fill out as part of a survey that was being conducted. During the analysis of the data, the PLS-SEM approach was utilized. When it came to the decision of whether or not students would continue to use the metaverse for educational purposes in the future, autonomy, and PU were significant indications. On the other hand, PEOU was not a significant factor in determining continuous intention. In addition, the research showed that autonomy and relatedness are believed to be important predictors of both PU and PEOU, whereas competence was shown to be an inconsequential predictor. Furthermore, according to the ANN model, relatedness, PEOU, autonomy, and PU are the most significant predictors, with a normalized value of 100%. The results that are obtained through the use of the PLS-SEM and ANN modes are identical.

This piece of work has a few shortcomings that need to be discussed and recognized. As a result of time constraints, the current study is cross-sectional, which means that it quantifies intentions at a certain point in time. The accumulation of experience causes individuals to develop new perspectives during the course of their lives. To this end, the longitudinal method is the most appropriate option to use. Only one private university was represented among the participants in the survey; there were no representatives from public universities. Therefore, it is not possible to generalize the findings of this study to other higher education institutions in Egypt. As a consequence of this, it is possible that future research will involve students from both public and private universities in order to contrast two distinct scenarios and generalize the findings. If various theories of technology acceptance are incorporated into future studies, it will be possible to gain a better understanding of the areas of interest that students have as well as the factors that influence their adoption of new technologies. It is possible that future research will take into account more key aspects, such as cognitive capacity, social relationships, and concerns about safety. The provision of connectivity can be of assistance in the development of virtual learning systems, particularly in light of the fact that a dependable and stable internet connection is quite important for metaverse technologies. The functionality of electronic devices may be negatively impacted if enough infrastructural support is not provided enough. The conclusion that can be drawn from this is that internet services must be of the best possible quality. When it comes to teaching, qualitative methodologies should be utilized in order to gain a deeper comprehension of the attitudes and behaviors of younger students. Take, for instance, the possibility of gaining more information through the use of participant interviews or focus groups.