Student Opinion Mining About Instructor Using Optimized Ensemble Machine Learning Model and Feature Fusion

Reviews given by students are an excellent source of information that can be used to achieve the educational goal of any institution. The information collected can be used to make administrative decisions, improve the teaching–learning process, improve student performance, and decrease course withdrawal. Processing student feedback is challenging due to the kind of language used by students and the large amount of data. The instructor is an important part of the university and plays an important role in the success of students and the university's reputation. Therefore, an efficient sentiment analysis model is required for processing student feedback to evaluate the instructor's performance. The feature extraction technique plays an important role in the performance of the sentiment analysis model. Recent studies have used word2vec/glove/n-gram/lexicon-based feature extraction techniques individually and not utilizing transformer-based contextual word representation techniques. For better feature representation, four types of features (transformer-based contextual features, word n-gram features, linguistic features, and lexicon-based features) are combined to form feature fusion. A soft voting ensemble model is applied considering the best three models from fifteen baseline models from three different categories (machine learning, deep learning, and transformer model). The weight parameters of each baseline model are optimized using a genetic algorithm. The robustness of the proposed model is evaluated on two datasets namely, the Vietnamese student feedback dataset and the dataset scrapped from RateMyProfessor.com. The developed model showed a 3% improvement in F1-Score over state-of-the-art models in the case of the Vietnamese student feedback dataset and a 7% improvement in the case of the RateMyProfessor dataset over baseline models. The ablation study showed that each component plays a role in the performance of the hybrid model.