Abstract
Despite the proliferation of online learning platforms, high student churn remains a significant obstacle for providers. This paper proposes a hybrid methodology to understand and predict student persistence ("stickiness"). By integrating a Particle Swarm Optimization (PSO)-enhanced Support Vector Machine (SVM) with Structural Equation Modeling (SEM), the study identifies a "multiple mediation effect." Findings reveal that student satisfaction does not merely impact persistence directly; rather, it operates through mediating variables such as content quality and performance expectations. The optimized SVM model demonstrates superior predictive accuracy for learning duration, providing a robust tool for platform providers to enhance user retention.
Introduction
The rapid expansion of digital education has highlighted a critical challenge: maintaining student engagement over time. While traditional research often focuses on direct satisfaction, this study addresses the "multiple mediation effect"—the phenomenon where primary drivers like satisfaction ripple through intermediate psychological variables to influence the final outcome of persistence.
The authors bridge the gap between predictive analytics and explanatory psychology. They aim to create a model that not only predicts how long a student will remain active on a platform but also explains the underlying mechanisms of that behavior, categorizing factors into platform-driven and student-driven variables.
Methodology
The research employs a dual-layered approach to analyze student behavior:
1. Predictive Modeling: PSO-SVM
To forecast learning duration, the authors developed an improved Support Vector Machine (SVM). The performance of an SVM is highly sensitive to the penalty factor ($Z$) and the kernel parameter ($\Psi$). To find the optimal configuration, the researchers utilized Particle Swarm Optimization (PSO), an evolutionary algorithm that mimics social foraging behaviors to search the parameter space efficiently.
2. Explanatory Modeling: SEM and Mediation Analysis
Structural Equation Modeling (SEM) was used to analyze the relationships between six core variables. These include endogenous factors (performance expectation, content quality, satisfaction degree) and exogenous factors (flow experience, perceived mobility, service quality).
The mediation analysis specifically quantified the "total effect" of satisfaction on persistence. The study utilized two primary datasets: 2,100 data points for learning duration to train the SVM, and survey data from approximately 152 students for the structural analysis shown in the path diagram above.
Results and Discussion
Enhanced Prediction Accuracy
The PSO-optimized SVM significantly outperformed standard SVMs and those optimized by other algorithms (e.g., Genetic Algorithms or Fruit Fly Optimization). In the [80, 100] minute learning duration range, the proposed model maintained an accuracy above 0.95. Visual analysis of absolute errors confirmed a marked reduction in prediction variance following PSO implementation.
The Mediation Mechanism
The most critical insight is the confirmation of a "partial mediating effect." Satisfaction improves a student's perception of content quality and their own performance expectations. These improved perceptions, in turn, are the primary drivers of long-term persistence. Furthermore, "Flow Experience"—the state of deep immersion—was found to be a prerequisite for high-quality user experience, suggesting that platform design must facilitate this psychological state to ensure "stickiness."
Conclusion
This research demonstrates that student persistence is a multi-faceted construct. By combining the predictive power of PSO-SVM with the explanatory depth of SEM, the study provides a comprehensive framework for understanding online learning behavior. While the current sample size for the structural model is a limitation, the findings offer clear directions for platform operators: focus on enhancing content quality and fostering "flow" to leverage the mediation effects of student satisfaction. Future work should explore these dynamics across diverse cultural and corporate training contexts.
Summary updated