Semantic analysis implementation using TF-IDF and RoBERTa in collaborative filtering

Abstract

Recommender systems (RSs) are used to provide recommendations that suit user preferences based on item features or ratings. In this research, we propose implementing TF-IDF and RoBERTa to learn and analyse keyword importance and semantic features in users reviews. TF-IDF is a statistical method that calculates the number of terms appearing in a document and calculates the importance of terms appearing in the document. RoBERTa is a robustly optimized BERT model that is trained more efficiently than BERT with better performance. These approaches produce semantic labels of positive, neutral and negative along with their probability values and are then evaluated using precision, recall, F1 score, accuracy and receiver operating characteristic (ROC) curve. The values are then normalized to 1-5 range before adjusting the original ratings. These new ratings are fed into user-based and item-based collaborative filtering systems and the recommendation results are evaluated using RMSE and MAE to compare the performance of all systems. Based on the experimental results, combining both TF-IDF and RoBERTa produces the highest accuracy for sentiment analysis and the lowest RMSE value in collaborative filtering.