Least Squares Approach to Locally Weighted Naive Bayes Method

This study proposes a new approach which calculates the weights of Locally Weighted Naive Bayes (LWNB) developed on Naive Bayes (NB) which is known with its simple structure. In this approach, a new equation is described by assigning a powered weight to each probabilistic factor in classic NB, and it is transformed to a linear form by using a simple assumption based on a logarithmic process, and then the weights are estimated by least squares technique. The success ratios are computed on two-class datasets from UCI database. The results show that LWNB with proposed approach is more successful than classic NB. In another analysis, it is determined that the class probability factor may sometimes damage the classification success. In addition, the effects of the attributes on the classification success are researched and according to the results the new approach is also suggested in the using as a feature selection technique of the pattern recognition problems.