A psychological approach to learning causal networks

Abstract

We examine the role of a common cognitive heuristic in unsupervised learning of Bayesian probability networks from data. Human beings perceive a larger association between causal than diagnostic relationships. This psychological principal can be used to orient the arcs within Bayesian networks by prohibiting the direction that is less predictive. The heuristic increased predictive accuracy by an average of 0.51 % percent, a small amount. It also increased total agreement between different network learning algorithms (Max Spanning Tree, Taboo, EQ, SopLeq, and Taboo Order) by 25 %. Prior to use of the heuristic, the multiple raters Kappa between the algorithms was 0.60 (95 % confidence interval, CI, from 0.53 to 0.67) indicating moderate agreement among the networks learned through different algorithms. After the use of the heuristic, the multiple raters Kappa was 0.85 (95 % CI from 0.78 to 0.92). There was a statistically significant increase in agreement between the five algorithms (alpha < 0.05). These data suggest that the heuristic increased agreement between networks learned through use of different algorithms, without loss of predictive accuracy. Additional research is needed to see if findings persist in other data sets and to explain why a heuristic used by humans could improve construct validity of mathematical algorithms.