Classifying YouTube Channels: a Practical System

This paper presents a framework for categorizing channels of videos in a thematic taxonomy with high precision and coverage. The proposed approach consists of three main steps. First, videos are annotated by semantic entities describing their central topics. Second, semantic entities are mapped to categories using a combination of classifiers. Last, the categorization of channels is obtained by combining the results of both previous steps. This framework has been deployed on the whole corpus of YouTube, in 8 languages, and used to build several user facing products. Beyond the description of the framework, this paper gives insight into practical aspects and experience: rationale from product requirements to the choice of the solution, spam filtering, human-based evaluations of the quality of the results, and measured metrics on the live site.

• Data Mining and Modeling