Meta learning for selection of best causal discovery algorithms.

Abstract

Selection of the best causal discovery algorithm for any new dataset is a difficult and time consuming process as it requires a researcher to have prior knowledge about a number of existing standard structure learning algorithms. During this research, we proposed a novel meta-learning approach to this problem. Meta-learning refers to learning about learning algorithms where different kinds of meta-data, such as properties of the learning problem, performance measures of different algorithms and patterns previously derived from the data are used to select the best or combine different learning algorithms to effectively solve a given learning problem. Several Bayesian networks in literature were manipulated, sampled to generate thousands of datasets, and specific features were extracted from each for meta-learning. Three standard structure learning algorithms were run on each of the generated datasets to discover the underlying causal networks and their performance was evaluated. With our new techniques, we were able to implement a tool for generating of many causal models and sampling many datasets from each model. We were able to determine the best algorithm or a combination of algorithms for specific datasets based on features extracted from them.