A K-NN associated fuzzy evidential reasoning classifier with adaptive neighbor selection

Abstract

We present a fuzzy evidential reasoning algorithm in light of the Dempster-Shafer evidence theory and the K-nearest neighbor algorithm for pattern classification. Given an input pattern to be classified, each of its K nearest neighbors is viewed as an evidence source, in terms of a fuzzy evidence structure. The distance between the input pattern and each of its K nearest neighbors is used for mass determination while the contextual information of the nearest neighbor in the training sample space is formulated by a fuzzy set in determining a fuzzy focal element. Therefore, pooling evidence provided by neighbors is realized by a fuzzy evidential reasoning, where feature selection is further considered through ranking and adaptive combination of neighbors. A fast implementation scheme of the fuzzy evidential reasoning is also developed. Experimental results of classifying multichannel remote sensing images have shown that the proposed approach outperforms the K-nearest neighbor (K-NN) algorithm [T.M. Cover et al. (1967)], the fuzzy K-nearest neighbor (F-KNN) algorithm [J.M. Keller et al. (1985)], the evidence-theoretic K-nearest neighbor (E-KNN) algorithm [T. Denoex (1995)], and the fuzzy extended version of E-KNN (FE-KNN) [L.M. Zouhal et al. (1997)], in terms of the classification accuracy and insensitivity to the number K of nearest neighbors.