Feature Selection for Multilabel Classification of Student Feedback Using Filter and Metaheuristic Methods

This study enhances multilabel classification of student feedback using both filter-based and metaheuristic feature selection methods. Student feedback offers crucial insights into educational services, covering areas like course quality and facility satisfaction. While previous research has focused mainly on sentiment analysis, this study categorizes multiple aspects of educational services in each feedback entry, providing a more comprehensive view for institutional improvement. The study uses BERT (Bidirectional Encoder Representations from Transformers) to transform feedback text into numerical vectors, leveraging BERT's ability to capture contextual meaning for accurate classification. A Random Forest Classifier, combined with a Multi Output Classifier, is used to manage the multilabel data structure. Filter-based methods tested include Variance Threshold, ANOVA, ReliefF, Mutual Information, Fisher Score, Pearson Correlation, Mean Absolute Difference, Dispersion Ratio, and Recursive Feature Elimination (RFE). Metaheuristic methods include Particle Swarm Optimization (PSO) and Grey Wolf Optimizer (GWO). Among filter-based methods, Variance Threshold achieved the highest accuracy 0.973, with Fisher Score and ReliefF maintaining high performance with fewer features. PSO demonstrated strong performance, achieving an accuracy of 0.960, with high precision 0.909 and an F1 score of 0.837, and surpassing GWO in precision. In conclusion, feature selection via filter-based or metaheuristic methods significantly improves classification accuracy, enhancing institutions' ability to analyze student feedback effectively and improve service quality.