Evidence Combination for Pattern Classification With Missing Data

Abstract

It remains a challenging problem to classify the incomplete patterns with randomly missing values. In some applications, it is difficult for us to collect complete attributes of target due to the complex sensing environment, and the observed patterns are all with more or less missing values. In order to well classify such incomplete patterns, we propose a new classification method based on evidence combination. Because the patterns are all considered with missing attribute values, it is hard to accurately estimate the missing values of object to classify. We propose to classify the object using each available attribute respectively. The K-nearest neighbors (K-NN) of object are found in training data space according to each attribute, and K Basic belief assignments (BBA) are constructed corresponding to the K-NN. The BBA reflects the degree of object belonging to each class. Then a two step evidence combination strategy is developed for combining the K BBA's. The BBA's associated with the same class are combined by classical DS rule at first. When the K-NN belong to different classes, the previous DS combination results associated with different classes may highly conflict, and they are further combined by PCR5 rule, which can well manage the high conflict via proper conflicting masses redistribution. We can classify the object using other attribute one by one in the similar way. When multiple classification results corresponding to different attributes are collected, a weighted evidence combination method is employed to fuse these results. The weighting factors of the results are optimized by minimizing an error criterion. The object is finally classified depending on this combination result. Experimental results with various data sets show the effectiveness of the proposed method comparing with other related methods.