A Causal Model Using Self-Organizing Maps

Understanding causality is very important for problem solving in many areas. However, conducting causal analysis for multivariate and nonlinear data, unlabeled in nature, still faces many problems with existing methods. Artificial Neural Networks have been developed for such data analyses and Backpropagation Network has been most used to learn the relationships between causes and effects. However, its supervised and black-boxed learning structure for labeled data often limits modeling and reasoning of uncertain, graded and fuzzy causality through the network. In this paper, an approach for analyzing such causality is proposed by networking Self-Organizing Maps handling unlabeled data. A new weighting scheme on connection weight vector similarity is developed to approximate conditional distributions of data in order to capture the indeterminate nature of causality. The experiments demonstrate that the method well approximates conditional output distributions for given inputs and allows estimating causal effects based on the similarity weight distributions.