Abstract

Critics of automatic causal discovery have claimed that Tetrad-style algorithms are inferior to domain experts at discovering causal structure from real scientific data and especially poor when applied to data that is highly mixed, either in a physical sense or in a mixtures of records sense. We compare a domain expert in geological spectroscopy head-to-head with a variety of machine algorithms on the task of predicting mineral class composition from visual to near infrared reflectance spectra. A simplified Tetrad algorithm outperforms all other machine algorithms tested and performs comparably to the domain expert. We conclude: that Tetrad algorithms can perform as well as human experts on this task, and that mixtures do not necessarily undermine the reliability of Tetrad algorithms on this task. This constitutes a counter-examples to the claim that machine algorithms are necessarily inferior to domain experts on tasks involving causal reasoning with real scientific data