12/07/2008uai 2008 cumulative distribution networks and the derivative-sum-product algorithm jim c....

20
12/07/2008 UAI 2008 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

Upload: prosper-jennings

Post on 12-Jan-2016

213 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Cumulative Distribution Networks and the Derivative-

Sum-Product Algorithm

Jim C. Huang and Brendan J. FreyProbabilistic and Statistical Inference Group, Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON, Canada

Page 2: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

• Problems where density models may be intractable

• e.g.: Modelling arbitrary dependencies

• e.g.: Modelling stochastic orderings

• Cumulative distribution network (CDN)

Motivation

e.g.: Predicting game outcomes in Halo 2

Page 3: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Cumulative distribution networks (CDNs)

• Graphical model of the cumulative distribution function (CDF)

• Example:

Page 4: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Positive convergence

Negative convergence

Monotonicity

Cumulative distribution functions

• Marginalization maximization

• Conditioning differentiation

Page 5: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Necessary/sufficient conditions on CDN

functions• Negative convergence (necessity and sufficiency):

• Positive convergence (sufficiency):

For each Xk, at least one neighboring function 0

All functions 1

Page 6: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 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

Page 7: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Marginal independence• Marginalization maximization

– e.g.: X is marginally independent of Y

Page 8: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 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

Page 9: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Check:

A toy example

Markov random fields

Required “Bayes net”

Page 10: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 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)

Page 11: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Derivative-sum-product

• In a CDN:

• In a factor graph:

Page 12: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Ranking in multiplayer gaming• e.g.: Halo 2 game with 7 players, 3 teams

Player skill functions

Player performance

Team performance

Given game outcomes, update player skills as a function of all player/team performances

Page 13: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Ranking in multiplayer gaming

= Local cumulative model linking team rank rn

with player performances xn

e.g.: Team 2 has rank 2

Page 14: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Ranking in multiplayer gaming

Enforce stochastic orderings between teams via h

= Pairwise model of team ranks rn,rn+1

Page 15: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

• CDN functions = Gaussian CDFs

• Skill updates:

• Prediction:

Ranking in multiplayer gaming

Page 16: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Results• Previous methods for ranking players:

– ELO (Elo, 1978)– TrueSkill (Graepel, Minka and Herbrich, 2006)

• After message-passing…

Page 17: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 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

Page 18: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 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– …

Page 19: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 2008

Thanks

• Questions?

Page 20: 12/07/2008UAI 2008 Cumulative Distribution Networks and the Derivative-Sum-Product Algorithm Jim C. Huang and Brendan J. Frey Probabilistic and Statistical

12/07/2008UAI 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