Discovery Systems Laboratory,.pdf

Algorithms for Large-Scale Local Causal Discovery and Feature Selection In the Presence Of Limited Sample Or Large Causal Neighbourhoods C.F. Aliferis and I. Tsamardinos Department of Biomedical Informatics, Discovery Systems Laboratory, Vanderbilt University Technical Report DSL-02-08 October 8th 2002 1 Introduction When the Causal Neighbourhood of a target node T is so large relative to the available sample then one cannot obtain statistically reliable tests of independence and association metrics when conditioning on all members of the causal neighbourhood. Thus both the growing and the shrinking phase of the algorithms in the IAMB family [Aliferis and Tsamardinos 2002a] will fail to give reliable results. Such situations arise often for example in text categorization, in studies with limited sampling (e.g., gene expression experiments), or when one wishes to learn the causal neighbourhood of a constellation of variables (e.g., in epidemiology when learning the causal neighbourhood of a “disease” that in effect summarizes the states of several variables some or all of which are also measured). We give here algorithms that circumvent the problem by trading off less sample (equivalently large causal neighbourhood when the sample is fixed) with computational complexity and with the possibility that false positives may be introduced in the estimated causal neighbourhood. We collectively call these algorithms S/LCN (Small Sample/Large Causal Neighbourhood). The basic structure for all such algorithms is identical, so we will first outline the main concept and see how different members of this family can be created by changes in the basic structure. We will also discuss how the same techniques we introduced for the IAMB family for chunking and parallelization can be directly applied in the S/LCN family to provide distributed processing/storage versions. The Main