An evidential K-nearest neighbor classification method with weighted attributes

The evidential K-nearest neighbor (EK-NN) method, which extends the classical K-nearest neighbour (K-NN) rule within the framework of evidence theory, has achieved wide applications in pattern classification for its better performance. In EK-NN, the similarity of test samples with the stored training ones are assessed via the Euclidean distance function, which treats all attributes with equal importance. However, in many situations, certain attributes are more discriminative, while others may be less irrelevant, so attributes should be weighted differently in distance function. In this paper, a new evidential K-nearest neighbor classification method with weighted attributes (WEK-NN) is proposed to overcome the limitations of EK-NN. In WEK-NN, the class-conditional weighted Euclidean distance function is developed to assess the similarity between two objects and both a heuristic rule and a parameter optimization procedure are designed to derive the attribute weights. Several experiments based on simulated and real data sets have been carried out to evaluate the performance of the WEK-NN method with respect to several classical K-NN approaches.