Context-aware movie recommendations with large language models

Abstract

The rapid growth of online streaming platforms has resulted in a vast information overload of movie content, presenting a challenge for users in their day to day lives. Effective recommendation systems are vital for improving user experience by providing personalizing content to individual users. This thesis introduces an advanced approach leveraging cutting- edge artificial intelligence technologies, employing ChatGPT along with text-embedding-3- large embedding model, integrated with the Weaviate vector database, to create dynamic recommendation model for movie recommendation. We have generated embedding vectors for 26,942 movies with additionally retrieved metadata, to build this personalized recommendation system. Two distinct recommendation models were developed and were used to compare their recommendation capabilities. The first model creates a centroid reference vector for each user based on their movie ratings of three stars or higher. This reference vector is then used to search for similar movies within the vector database using L2-squared distance, aiming to match user movie preferences. The second model refines this approach by adjusting the centroid vectors according to the weight of user ratings, allowing for a more complete reflection of user tastes. Comparative analysis on recall, precision, F1 score, and hit rate (HR) for the top 5 and top 10 recommendations shows Model 1 achieving a recall of 1.8% and 2.81%, precision of 17.01% and 14.03%, and F1 scores of 3.15% and 4.45% respectively. These metrics demonstrate its effectiveness, with HR in the top 5 and top 10 recommendations reaching 48.97% and 63.66%, respectively. This indicates that Model 1 not only identifies relevant movies with relatively better accuracy but also does so more consistently than Model 2 and is computationally efficient. The results of this study highlight the potential of using advanced AI models to significantly enhance the precision and user satisfaction of movie recommendation systems. This approach allows for a deeper understanding of user preferences and the ability to predict user needs, ultimately leading to a more engaging and satisfying viewing experience.