Movie Tag Prediction Using Multi-label Classification with BERT

Abstract

Recommendation systems are essential in optimizing user engagement on Over-the-Top (OTT) and Video-On-Demand (VOD) platforms. The conventional collaborative filtering approach, though effective, faces the cold-start challenge due to an undeveloped user base for new contents. To address this, many platforms turn to metadata-based recommendations, however, this method often struggles with content lacking rich metadata. This research introduces a novel solution that utilizes textual content—overviews, reviews, plots, and subtitles—to generate enhanced content descriptions. By integrating the BERT model, tailored for multi-label classification, and training it on the curated MovieLens Tag Genome 2021 dataset, we achieved dual outcomes: an improved similarity matrix using cosine similarity and a tag extraction system that aids in creating custom categories. This approach not only enhances content recommendation but also offers a feedback mechanism, promising a more enriched and personalised user experience. © 2024 Elsevier B.V., All rights reserved.