Making Recommendations for an Intelligent Video Channel

Collaborative filtering is a class of methods that analyze past user feedback to find relationship among users and items, or to predict a relationship between a user and a new item. Predicting user preferences is an important problem with primary applications in intelligent applications and recommender systems. Popular examples of such applications for audio and video are Netflix (movie recommendation system), and Last.fm/Pandora (Internet radio station that adjusts to user's music preferences), but there's still no individualized video channel that we are aware of. Building such video channel and developing an approach for making recommendations is the goal of this research. We developed and offline video player that works like a TV channel. Videos to play are downloaded from Youtube by a downloading service. Player gathers all user actions (stop, skip, scroll) and saves them in the database for analysis. We also gathered a database of information about more than 80 thousand of music videos posted on Youtube and more than two million users of Youtube. In addition to features from Youtube, we store features obtained from audio and video fingerprinting. Server-based recommendation system provides personalized recommendations for each user. This system can be extended to commercial to individualized TV channel as a next generation of 'video on demand'. We analyzed feasibility of many machine learning methods for rating prediction problem, and tested some of them on our data. We used machine learning tools provided by WEKA and LibSVM to experiment with different sets of features for rating, language, quality, and motion prediction. Our main contribution is discovery of simple features that allow predicting rating for video with high accuracy. These features are estimated ratings calculated from keywords, titles, and descriptions for videos mined from Youtube. Exact rating can be predicted with 50% accuracy on average, and with 85% average accuracy we can predict whether a user will like a video or not, based on 30 rated videos. Small size and simplicity of features allow for fast learning and prediction. Resulting user profile is also small because it is represented only by keywords weights. This is important if a channel has a large number of users. Our approach is important because it allows predicting rating for a new video. User-based approach cannot predict rating for a video until enough users rated it. For a video channel that can choose among millions of videos with new videos appearing every day, this is infeasible. Based on keywords we can immediately rate any Youtube video. To be able to provide personalized content to a first-time user, it's sufficient to obtain his ratings for a small initial set of videos with many keywords. The problem associated with this approach is that some videos that lack popular keywords may never be recommended. That could be overcome with alternative way to classify videos, e.g by predictiong music genre from audio features.