Portable Few-Shot Learning for Early Warning Systems in Small Private Online Courses: A CNN-Based Predictive Framework for Student Performance

Abstract

This study develops a portable early warning system designed to predict student academic performance in Small Private Online Courses (SPOCs) by leveraging few-shot learning techniques and convolutional neural networks (CNNs). Addressing the persistent challenges posed by limited sample sizes and the absence of face-to-face interactions in asynchronous online environments, the research explores whether small-sample behavioral data derived from multiple SPOCs can support reliable and transferable predictive models. The dataset comprises more than 4.4 million LMS log entries collected from four online courses sharing similar instructional designs and taught by a single instructor. After comprehensive preprocessing—encompassing feature extraction, weekly aggregation, and normalization—18 week-specific CNN models were trained to capture the temporal progression of student learning behaviors. The results indicate that meaningful prediction accuracy emerges by the fifth week, with performance exceeding 80% from week eight onward. Portability was further validated by applying the model to an additional course, where accuracy remained at or above 81%, confirming its robustness under consistent instructional conditions. The findings highlight the potential of few-shot learning to sustain predictive performance despite limited training samples, offering educators a viable foundation for timely interventions and institutional adoption of precision-driven academic support systems.