Advancing Personalized Recommendation Algorithms with RNN and Multimodal Features

Abstract

Traditional recommendation systems, employing collaborative filtering (CF) methods, recommend related videos for a given seed by considering how many users watched a particular candidate video immediately after the seed video. This approach primarily relies on aggregate user behavior and does not take into account the content of the videos. Current recommendation systems can recommend content across different media types, including text, images, audio, and video. This is achieved by pre-tagging various media files with different labels and then using search engines to find content based on user needs. However, with the evolution of video platforms, user demands for personalization have become stronger, and recommendation scenarios have become more complex. There is an increasing need for precision in video recommendations. To address the limitations, we focus on integrating multiple modalities into recommendation systems. The main contributions of this work are: We explore how to effectively integrate the various modalities to reflect user preferences and how they affect recommendation performance. We perform extensive experiments on public dataset to demonstrate the performance of our proposed model.