Emotion-Based Movie Recommendation System

Abstract

This study presents a novel approach for a movie recommendation system that uses the emotions of a user to recommend movies. To detect user emotions, the system uses both facial expressions and text analysis. To detect facial expressions, several types of pre-trained models were re-trained and evaluated using benchmark datasets (FER2013). The ResNet50 model which has the highest accuracy of 73% was selected as the final model. For text analysis, several classical machine learning models (SVM, RF, MNB) and deep learning models (LSTM, Bi-LSTM, BERT, BERT+CNN) were trained and evaluated for their effectiveness in classifying emotions (using ISEAR). The BERT+CNN model with an accuracy of 78% was ultimately chosen for its high accuracy and efficiency in handling textual data. Final emotion derived by applying soft voting ensemble technique to the results of facial expression model and the text analysis model. For making the recommendations, the study incorporated content-based and collaborative filtering techniques to recommend movies based on the users’ emotional state. Both methods were combined and adjusted based on the user’s emotional state, resulting in more personalized movie recommendations. To assess the efficiency of the proposed system, feedback was collected from ten users and analyzed. The final system received positive feedback from seven of the ten users. This indicated that the proposed system has the potential to enhance user experience by providing more personalized and relevant movie recommendations based on their emotional state.