Metodologia para recomendação de vídeos baseada em descritores de conteúdo visuais e textuais

Abstract

This work addresses the development of a new multimodal approach to deal with the cold-start problem in video recommender systems. Given a video document, which usually consists of a sequence of images, audio and related information (such as title, tags and description), video recommendation is formulated as finding a list of the most relevant videos to target users. The cold-start problem is caused when there is no information about the user (cold-start user) or the item (cold-start item) when recomending an item. More specifically, our approach considers two information modalities: (i) visual (color and facial information) and (ii) textual (tags, title, and description) information. To qualify the recommendation performed, we extend a Sparse linear method with side information (SSLIM), which takes into account the aforementioned information modalities. In this work, we propose an approach that exploits visual and textual information to properly describe videos in our recommender system. Experimental results obtained on a real dataset containing 207.154 videos from Youtube shows that we could improve up to 7% when comparing to a single modality model and up to 12% when compared to the baseline, demonstrating the effectiveness of the proposed approach and highlighting the usefulness of multimodal information when dealing with the cold-start problem.