A general dual-view framework for instance weighted naive Bayes

Instance weighting is an effective and flexible method to alleviate the attribute conditional independence assumption in naive Bayes (NB). However, existing instance weighting methods mainly focus on how to learn a specific weight for each instance, ignoring the limitation of the original view. In this study, we argue that real-world applications are rather complicated, and it is sub-optimal to learn instance weights only using the original view. Based on this premise, we propose a novel general framework called dual-view instance weighted naive Bayes (DIWNB). In DIWNB, we first construct multiple K-nearest neighbor (KNN) classifiers and select those with the lowest error rate to classify each training instance in turn to build the generated view. Next, we learn a specific weight for each training instance, and build an instance weighted NB model in each view. Finally, we weightedly fuse the class-membership probabilities of dual views to predict the class label for each test instance. To construct the generated view, we design a hard label approach and a soft label approach, and thus two different versions are created, which we denote as DIWNB${}^H$ and DIWNB${}^S$, respectively. Experimental results on 60 benchmark and 2 real-world datasets demonstrate the effectiveness of DIWNB.