Ensembled Identification for Problematic Student Based on Multi-perspective Analysis Using College Students’ Behavioral Data

Psychological health of college students has become one of the most critical problems in current higher education. Accordingly, the identification of problematic college students has attracted general concerns from the universities and society. But due to complex influencing factors of detection task and extremely imbalanced distribution of data used to train detection model, the existing detection approaches base upon a single source of college students’ behavioral data cannot be used to identify problematic college students effectively. In this paper, we regard the issue of problematic college student detection as a binary classification task, and propose an ensembled identification framework for problematic student (EIPS) based on multi-perspective analysis using college students’ behavioral data. Through introducing multi-head self-attention mechanism into training binary classification model, we can capture differentiated influences of distinct features on the classification task. Further, we utilize an ensemble framework to enhance classification performance by incorporating with the outputs of multiple base classifiers to obtain the final classification result. Finally, extensive comparative experiments on real data sets demonstrate that EIPS significantly outperforms the state of-the-art methods.