Regression Can Build Predictive Causal Models

Covariance information can help an algorithm search for predictive causal models and estimate the strengths of causal relationships. This information should not be discarded after conditional independence constraints are identi ed, as is usual in contemporary causal induction algorithms. Our fbd algorithm combines covariance information with an e ective heuristic to build predictive causal models. We demonstrate that fbd is accurate and e cient. In one experiment we assess fbd's ability to nd the best predictors for variables; in another we compare its performance, using many measures, with Pearl and Verma's ic algorithm. And although fbd is based on multiple linear regression, we cite evidence that it performs well on problems that are very di cult for regression algorithms. This research is supported by ARPA under contract F30602-93-C-0100, and by a NASA GSRP Training Grant, #NGT-70358.