Causal discovery from changes.pdf

Causal Discovery from Changes Jin Tian and Judea PearlCognitive Systems LaboratoryComputer Science DepartmentUniversity of California, Los Angeles, CA 90024fjtian, judea g@AbstractWe propose a new method of discoveringcausal structures, based on the detectionof local, spontaneous changes in the un-derlying data-generating model. We ana-lyze the classes of structures that are equiv-alent relative to a stream of distributionsproduced by local changes, and devise algo-rithms that output graphical representationsof these equivalence classes. We present ex-perimental results, using simulated data, andexamine the errors associated with detectionof changes and recovery of structures.1 IntroductionIn recent years, several graph-based algorithmshave been developed for the purpose of infer-ring causal structures from empirical data. Someare based on detecting patterns of conditionalindependence relationships [Pearl and Verma, 1991,Spirtes et al., 1993], and some are based on Bayesianapproaches [Heckerman et al., 1997, Cooper, 1999].These discovery methods assume static environment,that is, a time-invariant distribution and a time-invariant data-generating model, while attempting toinfer structures that encode dynamic aspects of theenvironment, for example, how probabilities wouldchange as a result of interventions. This transition,from static to dynamic information, constitutes a ma-jor inferential leap, and is severely limited by the inher-ent indistinguishability (or equivalence) relation thatgoverns Bayesian networks [Verma and Pearl, 1990].One way of overcoming this basic limitation is toaugment the data with partial causal knowledge, ifsuch is available. [Spirtes et al., 1993], for exam-This is part-I of a two-part paper submitted to UAI-01. Part-II is entitled: Causal Discovery from Changes: aBayesian Approach. ple, discussed the use of experimental data to iden-tify causal relationships. [Cooper and Yoo, 1999] dis-cussed a Bayesian method of causal discovery from ami