Locally weighted learning: how and when does it work in Bayesian networks?

Abstract

Bayesian network (BN), a simple graphical notation for conditional independence assertions, is promised to represent the probabilistic relationships between diseases and symptoms. Learning the structure of a Bayesian network classifier (BNC) encodes conditional independence assumption between attributes, which may deteriorate the classification performance. One major approach to mitigate the BNC’s primary weakness (the attributes independence assumption) is the locally weighted approach. And this type of approach has been proved to achieve good performance for naive Bayes, a BNC with simple structure. However, we do not know whether or how effective it works for improving the performance of the complex BNC. In this paper, we first do a survey on the complex structure models for BNCs and their improvements, then carry out a systematically experimental analysis to investigate the effectiveness of locally weighted method for complex BNCs, e.g., tree-augmented naive Bayes (TAN), averaged one-dependence estimators AODE and hidden naive Bayes (HNB), measured by classification accuracy (ACC) and the area under the ROC curve ranking (AUC). Experiments and comparisons on 36 benchmark data sets collected from University of California, Irvine (UCI) in Weka system demonstrate that locally weighting technologies just slightly outperforms unweighted complex BNCs on ACC and AUC. In other words, although locally weighting could significantly improve the performance of NB (a BNC with simple structure), it could not work well on BNCs with complex structures. This is because the performance improvements of BNCs are attributed to their structures not the locally weighting.

Shirui Pan

Professor and ARC Future Fellow

My research interests include data mining, machine learning, and graph analysis.