Modification of belief in evidential causal networks

A data mining tool which addresses both the discovery and update of causal networks hidden in database systems, based on Evidence Theory, which offers important advantages over conventional Bayesian methods for the applications envisaged.

Abstract

This paper introduces a new evidential approach for the updating of causal networks which is to be added to an existing general data mining system prototype—the Mining Kernel System (MKS). We present a data mining tool which addresses both the discovery and update of causal networks hidden in database systems. It contributes to the discovery of knowledge which links rules—knowledge which would normally be considered domain knowledge (to be elicited from domain experts). We used different methods for generating networks such as our heuristic algorithm (HNG), which is briefly discussed in this paper. Evaluation of such knowledge presents difficulties but some anecdotal appraisal is presented here in the form of a simple case study. Applications of this prototype with its new causal updating supplement are under way. Our approach is based on Evidence Theory and offers important advantages over conventional Bayesian methods for the applications envisaged. These approaches allow certainty levels of rules in causal networks to be kept up to date. When a causal network has been discovered, any subsequent new evidence may be fed into the model. After updating the belief function for any node the complete network is updated through communication between neighbouring nodes.