Computational Thinking (CT) is a core competency in computer science education; however, its accurate assessment remains a persistent challenge due to the complex and dynamic nature of the programming processes of learners. The main problem is that current assessment approaches fail to fully capture how students develop CT skills over time, leading to incomplete or biased evaluation outcomes. The key gap in existing research is that most studies rely on either sequential behavioral models or static feature-based models, with limited integration of both temporal learning patterns and aggregated learner characteristics. This study proposes a hybrid deep learning framework that integrates long short-term memory networks and feed-forward neural networks to jointly model sequential code submission behavior and static learner attributes to address this gap. The model is evaluated using a large-scale programming dataset comprising 10,532 students and the following: 65,280 submission sequences. A five-fold stratified cross-validation approach is adopted to ensure the robustness and generalizability of the results. The findings show that the proposed hybrid model achieves an accuracy of 97.0%, an F1-score of 96.9%, and an ROC–AUC of 0.98, outperforming the baseline models, including the standalone LSTM, FFNN, and decision tree classifiers. Statistical significance testing confirmed that these improvements were not due to random variation (p < 0.05). This study contributes to educational data mining by demonstrating that integrating temporal learning dynamics with static learner features produces a more comprehensive and accurate assessment of computational thinking. It also provides empirical evidence supporting hybrid DLarchitectures for theory-informed and data-driven evaluation of programming skill development
Dumre et al. (Wed,) studied this question.