Tag suggestion on youtube by personalizing content-based auto-annotation

Abstract

We address the challenge of tag recommendation for web video clips on portals such as YouTube. In a quantitative study on 23,000 YouTube videos, we first evaluate different tag suggestion strategies employing user profiling (using tags from the user's upload history) as well as social signals (the channels a user subscribed to) and content analysis. Our results confirm earlier findings that --~at least when employing users' original tags as ground truth~-- a history-based approach outperforms other techniques. Second, we suggest a novel approach that integrates the strengths of history-based tag suggestion with a content matching crowd-sourced from a large repository of user generated videos. Our approach performs a visual similarity matching and merges neighbors found in a large-scale reference dataset of user-tagged content with others from the user's personal history. This way, signals gained by crowd-sourcing can help to disambiguate tag suggestions, for example in cases of heterogeneous user interest profiles or non-existing user history. Our quantitative experiments indicate that such a personalized tag transfer gives strong improvements over a standard content matching, and moderate ones over a content-free history-based ranking.