EXPLOITING CAUSAL INDEPENDENCE IN BAYESIAN NETWORK INFERENCE

A new method is proposed for exploiting causal independencies in exact Bayesian network inference. A Bayesian network can be viewed as repre senting a factorization of a joint probability into the multiplication of a set of conditional probabilities. We present a notion of causal independence that enables one to further factorize the conditional pro babilities into a combination of even smaller factors and consequently obtain a finer-grain factorization of the joint probability. The new formulation of causal independence lets us specify the conditional pro bability of a variable given its parents in terms of an associative an d commutative operator, such as ''or'', ''sum'' or ''max'', on the con tribution of each parent. We, start with a simple algorithm VE for Bay esian network inference that, given evidence and a query variable, use s the factorization to find the posterior distribution of the query. W e show how this algorithm can be extended to exploit causal independen ce. Empirical studies, based on the CPCS networks for medical diagnosi s, show that this method is more efficient than previous methods and a llows for inference in larger networks than previous algorithms.