Adaptive Deep Learning Framework for Emotion Classification in Student Data Using Transformer Models and Advanced Evaluation Metrics

Abstract

Students' growing usage of social media has produced large datasets with insightful information about their emotional expressions and behavioral patterns. These databases offer opportunities to forecast success in online learning platforms through tailored and flexible interventions. With a focus on strong preprocessing, feature selection, and classification techniques, this work presents a sophisticated deep learning framework for emotion classification. This approach shifts to classification goals, employing metrics like accuracy, precision, recall, and F1-score for efficient evaluation, in contrast to earlier research that concentrated on prediction tasks using metrics like MSE, RMSE, and MAPE. Important preprocessing methods, such as Box-Cox transformation, are used to normalize data and improve model stability. The extraction of significant features is ensured by adaptive feature selection based on a Tversky index based on R\u00e9nyi entropy. Deep contextual linkages in text data are captured via transformer-based designs, such as BERT. Six emotion categories—Sadness, Joy, Love, Anger, Fear, and Surprise—found in a tagged student social media dataset are used to validate the framework, showing notable gains in classification performance. The findings highlight the method's potential for use in sentiment analysis, online learning platform performance prediction, and student well-being monitoring, ultimately facilitating data-driven educational decision-making.