We consider the problem of algorithmically recommending items to users on aYahoo! front page module. Our approach is based on a novel multilevelhierarchical model that we refer to as a User Profile Model with GraphicalLasso (UPG). The UPG provides a personalized recommendation to users bysimultaneously incorporating both user covariates and historical userinteractions with items in a model based way. In fact, we build a per-itemregression model based on a rich set of user covariates and estimate individualuser affinity to items by introducing a latent random vector for each user. Thevector random effects are assumed to be drawn from a prior with a precisionmatrix that measures residual partial associations among items. To ensurebetter estimates of a precision matrix in high-dimensions, the matrix elementsare constrained through a Lasso penalty. Our model is fitted through apenalized-quasi likelihood procedure coupled with a scalable EM algorithm. Weemploy several computational strategies like multi-threading, conjugategradients and heavily exploit problem structure to scale our computations inthe E-step. For the M-step we take recourse to a scalable variant of theGraphical Lasso algorithm for covariance selection. Through extensiveexperiments on a new data set obtained from Yahoo! front page and a benchmarkdata set from a movie recommender application, we show that our UPG modelsignificantly improves performance compared to several state-of-the-art methodsin the literature, especially those based on a bilinear random effects model(BIRE). In particular, we show that the gains of UPG are significant comparedto BIRE when the number of users is large and the number of items to selectfrom is small. For large item sets and relatively small user sets the resultsof UPG and BIRE are comparable. The UPG leads to faster model building andproduces outputs which are interpretable.
展开▼
