Search based Video Recommendations

In this paper, we present a search powered approach we have used in building a Video Recommendations Engine for Yahoo hosted videos and Yahoo Video Search. The aim is to increase user engagement by recommending related videos and hence increase revenue by being able to show more advertisements as the user keeps consuming more videos. This system accepts an input context which provides information about the user and the video consumed and returns a set of related videos as recommended. We look at this problem as a multi-faceted problem since the intent of the user at a particular point in time cannot be known deterministically. So we generate the candidate set of recommendations using an ensemble of algorithms and available search signals. We discuss these algorithms and mechanisms for retrieving related videos in details along with an explore-exploit strategy to learn a near optimal ranking of the candidate recommendations, and provide the performance results. This system has been able to increase the number of video plays at Yahoo by 66%.