Matrix Factorization for Movie Recommendations – CME 510 скачать в хорошем качестве

Matrix Factorization for Movie Recommendations – CME 510

Matrix Factorization for Movie Recommendations Harald Steck, Netflix hsteck@netflix.com http://videolectures.net/harald_steck/ The Netflix recommender system for movies and TV shows is comprised of an ensemble of models. The talk will focus on matrix factorization models. Users' feedback data (eg, played or rated titles) can be represented in a matrix involving users and movies/TV shows. Such a matrix has several interesting properties: (1) it is sparse (ie each user rated only a small number of titles), (2) it is tall and thin (ie there are many more users than titles), and (3) there are various selection biases in the data. The latter means that there is information in which entries are present in the sparse matrix (besides the information in the entries' values). An example of a selection bias is that a user tends to rate items that they like or know, resulting in an under-representation of low rating values in the data. Another example is that users tend to rate movies with similar release-years together. I will discuss different matrix factorization models tailored to these properties of the data. The models are optimized by stochastic gradient descent toward a personalized ranking of the movies for each user, rather than toward predicting missing entries in the matrix.