Design of a Hybrid System Architecture for Predicting Students' Computational Proficiency Using Behavioral and Performance Data

The assessment of students’ programming is central to the education of computer science and is also difficult for students’ process and dynamically oriented programming skills. This study aims to create a hybrid FFNN LSTM structure to assess the CT proficiency of students from large-scale programming activities. The structure learns performance features and programming sequences from the dataset created by Azcona et al. in 2020, which includes 10,532 students and more than 65,280 code submission sequences. A stratified 5-fold cross-validation was used to evaluate the model. The results show that the hybrid model achieved more than 97.1% in average accuracy, 96.9% in average f1-score, and 0.98 in average ROC-AUC of 0.98, which outperformed the standalone FFNN LSTM and Decision Tree benchmark models. The significance of the performance improvement was statistically proven (p <0.05) through a paired t-test and one-way analysis of variance. The results indicate that by incorporating both static and dynamic features of learners, they can obtain a more accurate assessment of one’s CT proficiency, as well as the potential of hybrid models in education analytics and automated data-driven-teaching strategies.