CoFiRank : Maximum Margin Matrix Factorization for Collaborative Ranking Markus Weimer, Alexandros Karatzoglou, Quoc Viet Le and Alex Smola NIPS’07 Idea • Maximum Margin Matrix Factorization • Structured Estimation for Ranking • Bundle Method Solver Collaborative Filtering • Based on partial observed matrix to predict unobserved entries Matrix Factorization • Low Rank Approximation • SVD for fully observed Y • Non-convex Maximum Margin Matrix Factorization • Trace norm+Hinge loss: Convex • Semi-Definite Programming Regularized Matrix Factorization • Formulation minimize U,V Alternating optimizing L(UV , Ytrain) trUU 2 T trVV T • Probabilistic Matrix Factorization (PMF) L(UV , Ytrain) UV Ytrain 2 2 • CoFiRank ( , Ytrain) 1 NDCG( , Ytrain) • Linear Convex Upper Bound L( f , Ytrain) : max ( , Ytrain) c, f f Non-Convex Solved by linear programming How to Compute Loss? • Linear Convex Upper Bound L( f , Ytrain) : max ( , Ytrain) c, f f • Solved by Linear Programming min X C i j i, j X i, j Can this explain in simple way? Useful Links • CoFiRank http://www.cofirank.org • MMMF http://ttic.uchicago.edu/~nati/mmmf/ • MF http://helikoid.si/mf/index.html Famous Researchers in Optimization • Yurii Nesterov – “Introductory Lectures on Convex Optimization: A Basic Course” http://www.core.ucl.ac.be/~nesterov/ • Arkadi Nemirovski – “Efficient methods in convex programming” http://www2.isye.gatech.edu/~nemirovs/ • Stephen P. Boyd – “Convex Optimization” http://www.stanford.edu/~boyd/ • Stephen J. Wright – “Numerical Optimization” http://pages.cs.wisc.edu/~swright/ • Dimitri Bertsekas – “Nonlinear Programming” http://web.mit.edu/dimitrib/www/home.html Questions? Normalized Discounted Cumulative Gain (NDCG) How to set c? ( , f ) : c, f • ci is set decreasing, is maximized with respect toπ for argsort(f) • ci =(i+1)-0.25 Convex Upper Bound • Linear Convex Upper Bound L( f , Ytrain) : max ( , Ytrain) c, f f L( f ,Ytrain) ( * ,Ytrain) c, f f ( * ,Ytrain) sort (a), sort (b) a, b Bundle Method • General convex optimization solver with tight convergence bound O(1/) Bundle Method for CoFiRank