Posted by Chris Newell on , last updated
There are so many programmes to choose from these days it can be hard to decide what to watch. Personalised recommender systems can help but usually require knowledge of your previous viewing history or preferences. Unfortunately, some users may be new to our services, may not log-in or may not want their viewing history to be tracked. For these reasons we have been exploring the concept of a "standalone recommender" that does not depend on historical viewing records.
In this first experiment we have created a standalone recommender prototype called Sibyl which uses a novel drag and drop interface. This provides a fast and easy way for the user to express their viewing preferences by dragging programmes into “Like” or “Dislike” boxes. The user can also influence the recommender by dragging individual programmes out of the recommendations into the same boxes.
Because of the highly interactive nature of the interface we have chosen to use our client-side recommender rather than a server-based system. This makes the interface very responsive and also helps to maintain privacy as the user’s “Like” and “Dislike” choices are never sent to our servers. The K-Nearest Neighbor model used by the recommender is based on metadata available from BBC Programmes and is built at regular intervals on the server using the Java version of the MyMediaLite Recommender System Library.
In future experiments we hope to explore the benefits of storing the user’s viewing history or preferences locally within the browser to help improve the recommendations.