A multi-dimensional hybrid stochastic model for early and interpretable blockage detection in programming education

Abstract

Abstract Detecting student difficulties in programming education is a critical challenge for instructors, a task further complicated by the rise of AI coding assistants that can obscure the learning process. While fine-grained interaction logs from development environments offer rich behavioral data, existing analytical models often focus on a single dimension, limiting their predictive power and interpretability. This paper proposes a multi-dimensional hybrid stochastic model that integrates three complementary analytical layers to provide a more holistic understanding of student behavior. The model combines: (1) a Markov Chain to capture the behavioral dimension of workflow transitions; (2) a Hidden Markov Model (HMM) to infer the latent cognitive dimension, identifying states such as progression, hesitation, and blockage; and (3) a Recurrent Neural Network (RNN) with an attention mechanism to model the sequential dimension by capturing long-range temporal dependencies. We validated our approach on a dataset of interaction traces collected from 70 first-year computer science students. The hybrid model achieved a macro-averaged F1-score of 88.1\% across four distinct cognitive states (Progressing, Hesitating, Blocked, Confused), significantly outperforming all baseline and ablation models as confirmed by a Friedman test followed by a Nemenyi post-hoc analysis (p < .001). For the critical task of blockage detection, the model obtained a precision of 92.5\% and a recall of 90.0\%. By providing both high-level cognitive state labels and low-level attention-based explanations, our framework bridges the gap between predictive accuracy and pedagogical interpretability. The findings demonstrate that this integrated, multi-dimensional approach offers a robust and effective foundation for developing real-time, evidence-based support systems for programming education.