Abstract:

Technology adoption models in education have traditionally focused on explanatory analysis using Partial Least Squares Structural Equation Modeling (PLS-SEM), providing insights into the average causal relationships among determinants of technology use. PLS-SEM assesses and quantifies linear causal relationships that can be extrapolated from the sample to the population, thereby supporting macro-level decision-making. However, several researchers have recently emphasized the need to strengthen the predictive capabilities of these models to provide more personalized and actionable recommendations regarding the use of educational technologies. Such an approach can uncover nonlinear relationships that conventional PLS-SEM analyses do not reveal. Given that this line of inquiry has not been extensively explored in the field of education, the objective of this research is to integrate out-of-sample prediction using machine learning (ML) techniques into the PLS-SEM evaluation of learning management system (LMS) adoption, based on the unified theory of acceptance and use of technology (UTAUT). Three ML algorithms were employed: XGBoost, multilayer neural network (MNN), and support vector regression (SVR). These algorithms were selected for their strong predictive performance and the diversity of their modeling logics.

An online questionnaire was administered to university students enrolled in online courses supported by a proprietary LMS. A total of 477 valid responses were analyzed using SmartPLS v4.1.1.6. The analysis assessed the measurement model (item loadings, reliability, convergent and discriminant validities) and the structural model (collinearity, path coefficients, t-values, p-values, coefficient of determination, model fit, and effect sizes). Hypothesis testing revealed that performance expectancy (PE) is the most significant predictor of behavioral intention (BI). Social influence (SI) and facilitating conditions (FC) also exert meaningful effects on BI, resulting in a coefficient of determination R² of 0.239. The PLSpredict results indicated a predictive accuracy Q² of 0.213, a root mean squared error (RMSE) of 0.989, and a mean absolute error (MAE) of 0.560. In comparison, the machine learning models achieved substantially stronger predictive outcomes. The MNN model obtained the highest R² (0.381) and Q² (0.355), alongside the lowest RMSE (0.710) and the second-lowest MAE (0.491), underscoring its ability to capture complex, nonlinear patterns within the dataset. These findings indicate that nonlinear function estimators outperform traditional regression-based and linear prediction approaches for modeling behavioral intention in educational contexts.

This research contributes to the existing literature on technology adoption in education by revealing the complex nonlinear relationships among the key determinants of students’ intention to use LMS platforms. While traditional PLS-SEM analyses provide valuable average-level insights, they do not pinpoint the individual learners who may require immediate intervention regarding LMS adoption. By integrating predictive analytics, this study offers two contributions: a deeper theoretical understanding of LMS adoption patterns by demonstrating how ML models can complement PLS-SEM, and practical insights enabling university administrators to recognize the critical factors affecting the intention of technology use and to intervene with students at risk of early disengagement.

Keywords:

LMS, UTAUT, PLS-SEM, Machine Learning, Multilayer neural network.