Intervention Learning of Local Causal Structure Based on Sensitivity Analysis

Abstract

As intervened edges are difficult to be determined when intervention method is used for learning the causal relationships of probability model, an active learning method (Structural Intervention Learning of Sensitivity Analysis –SILSA Algorithm) is proposed. SILSA algorithm obtains original network structure based on k2 algorithm, then uses junction tree algorithm to decompose original networks structure and takes local intervention learning in every clique of junction tree, which can decrease the searching extension of intervened edges. Causal Bayesian networks can be learned by Edge-based Interventions when intervened edges are selected. In order to get appropriate intervened edge, sensitivity analysis is used to select the important edge in SILSA algorithm. The efficient of selecting intervened edge is improved. Experimental results show that the effectiveness and performance of SILSA algorithm are better than intervened edges with choosing randomly and passive learning method.