Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical Inference Group, Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON, Canada UAI 2008 12/07/2008 Motivation • Problems where density models may be intractable • e.g.: Modelling arbitrary dependencies e.g.: Predicting game outcomes in Halo 2 • e.g.: Modelling stochastic orderings • Cumulative distribution network (CDN) UAI 2008 12/07/2008 Cumulative distribution networks (CDNs) • Graphical model of the cumulative distribution function (CDF) • Example: UAI 2008 12/07/2008 Cumulative distribution functions Negative convergence Positive convergence Monotonicity • Marginalization maximization • Conditioning differentiation UAI 2008 12/07/2008 Necessary/sufficient conditions on CDN functions • Negative convergence (necessity and sufficiency): For each Xk, at least one neighboring function 0 • Positive convergence (sufficiency): All functions 1 UAI 2008 12/07/2008 Necessary/sufficient conditions on CDN functions • Monotonicity lemma (sufficiency): All functions monotonically non-decreasing… Sufficient condition for a valid joint CDF: Each CDN function can be a CDF of its arguments UAI 2008 12/07/2008 Marginal independence • Marginalization maximization – e.g.: X is marginally independent of Y UAI 2008 12/07/2008 Conditional independence • Conditioning differentiation – e.g.: X and Y are conditionally dependent given Z – e.g.: X and Y are conditionally independent given Z • Conditional independence No paths contain observed variables UAI 2008 12/07/2008 A toy example Required “Bayes net” Check: Markov random fields UAI 2008 12/07/2008 Inference by message passing • Conditioning differentiation • Replace sum in sum-product with differentiation … • Recursively apply product rule via message-passing with messages , • Derivative-Sum-Product (DSP) UAI 2008 12/07/2008 Derivative-sum-product • In a CDN: • In a factor graph: UAI 2008 12/07/2008 Ranking in multiplayer gaming • e.g.: Halo 2 game with 7 players, 3 teams Player skill functions Player performanc e Team performanc e Given game outcomes, update player skills as a function of all player/team performances UAI 2008 12/07/2008 Ranking in multiplayer gaming = Local cumulative model linking team rank rn with player performances xn e.g.: Team 2 has rank 2 UAI 2008 12/07/2008 Ranking in multiplayer gaming = Pairwise model of team ranks rn,rn+1 Enforce stochastic orderings between teams via h UAI 2008 12/07/2008 Ranking in multiplayer gaming • CDN functions = Gaussian CDFs • Skill updates: • Prediction: UAI 2008 12/07/2008 Results • Previous methods for ranking players: – ELO (Elo, 1978) – TrueSkill (Graepel, Minka and Herbrich, 2006) • After message-passing… UAI 2008 12/07/2008 Summary • The CDN as a graphical model for CDFs • Unique conditional independence structure • Marginalization maximization • Global normalization can be enforced locally • Conditioning differentiation • Efficient inference with Derivative-Sum-Product • Application to Halo 2 Beta Dataset UAI 2008 12/07/2008 Discussion • Need to be careful when applying to ordinal discrete variables… • Principled method for learning CDNs • Variational principle? (loopy DSP seems to work well) • Future applications to – Hypothesis testing – Document retrieval – Collaborative filtering – Biological sequence search –… UAI 2008 12/07/2008 Thanks • Questions? UAI 2008 12/07/2008 Interpretation of skill updates • For any given player let denote the outcomes of games he/she has played previously • Then the skill function corresponds to UAI 2008 12/07/2008