Case-Based Bayesian Network Classifiers

We propose a new approach for learning Bayesian classifiers from data. Although it relies on traditional Bayesian network (BN) learning algorithms, the effectiveness of our approach lies in its ability to organize and structure the data in such a way that allows us to represent the domain knowledge more accurately than possible in traditional BNs. We use clustering to partition the data into meaningful patterns, where each pattern is characterized and discriminated from other patterns by an index. These patterns decompose the domain knowledge into different components with each component defined by the context found in its index. Each component can then be represented by a local BN. We argue that this representation is more expressive than traditional BNs in that it can represent domain dependency assertions more precisely and relevantly. Our empirical evaluations show that using our proposed approach to learning classifiers results in improved classification accuracy.