Output details
11 - Computer Science and Informatics
University College London
Matchbox: large scale online bayesian recommendations.
<16> This paper presents Matchbox, a new probabilistic model and inference algorithm for recommender systems uniquely combining collaborative and content based recommendations. Instead of representing users and items as mere IDs, the model uses rich feature vectors to represent their meta-data. The rigorous formulation in terms of a fully probabilistic model makes it possible to mitigate the so-called cold-start problem and to track uncertainty associated with recommendations. The Matchbox algorithm is subject to US patent 12253854 and was the basis for the recently launched recommendation engine of Microsoft’s Xbox Live online gaming service with more than 40 million members.