Genre Effect Toward Developing a Multi-Modal Movie Recommendation System in Indian Setting
Prabir Mondal, Pulkit Kapoor, Siddharth Singh, Sriparna Saha, Jyoti Prakash Singh, Amit Kumar Singh · IEEE Transactions on Consumer Electronics · 2023
The Recommendation System (RS) does not require any introduction nowadays and its role in various platforms as an information filtering agent is very crucial. The personalized movie RS recommends the appropriate movies to the user after analyzing his preferences. Most of the recent studies in this RS basically consider the rating values of the user-movie pairs to evaluate user choice. However, our literature survey found that there is an unexplored area where the user feedback is in ordinal value and the movies are from the Indian regional language based. This paper has presented a multi-modal RS for the Hindi movie dataset where the user preferences over movies are from three different classes and these are, "Like", "Neutral" and "Dislike/ Not watched". The proposed model has employed the utilities of a multi-head cross-attention mechanism and as an input, it has used the extracted audio-video information from the movie trailers. Here the information on Hindi movies and their users has been taken from the Flickscore Dataset. The study evaluates a classification model’s effectiveness in two critical aspects: (i) the introduction of a novel metric called "GenreLike-Score (GL-score)," which aligns user genre preferences with movie genres, and (ii) the exploration of various audio-video embeddings. We have done thorough ablation studies using different combinations of these two factors and finally concluded how the GL-score is effective in predicting proper preference for user-movie pair. Besides that, different ways of keyframe extraction techniques as well as their different embedding generation processes have been evaluated in the ablation studies. Our classification model performs well when GL-score is considered and proper audio-video embeddings are determined for the studies.