Causal structure learning of highly context-dependent data

Résumé—This research was triggered by the analysis of performance data of compilers which is highly context-dependent. The output of the compiler and as such the performance of the compiled program varies a lot for different values and ranges of its categorical and continuous parameters. Context-dependence disrupts the functioning of statistical tests for identifying conditional independencies and causes the failure of independence-based causal structure learning algorithms. Test results for state-of-theart independence tests are presented. This paper then discusses a new approach for dealing with context-dependent data. First, the different contexts are learned using the M5P regression tree learning algorithm. Secondly, for each context a separate causal model is learned which results in a Bayesian multinet. Applied on the performance data, correct causal models are learned.