Pattern Classification using Rectified Nearest Feature Line Segment Hao Du and Yan Qiu Chen ⋆ Department of Computer Science and Engineering, School of Information Science and Engineering, Fudan University, Shanghai 200433, China Abstract. This paper proposes a new classification method termed Rec- tified Nearest Feature Line Segment (RNFLS). It overcomes the draw- backs of the original Nearest Feature Line (NFL) classifier and possesses a novel property that centralizes the probability density of the initial sample distribution, which significantly enhances the classification abil- ity. Another remarkable merit is that RNFLS is applicable to complex problems such as two-spirals, which the original NFL cannot deal with properly. Experimental comparisons with NFL, NN(Nearest Neighbor), k-NN and NNL (Nearest Neighbor Line) using artificial and real-world datasets demonstrate that RNFLS offers the best performance. 1 Introduction Nearest Feature Line (NFL) [1], a newly developed nonparametric pattern clas- sification method, has recently received considerable attention. It attempts to enhance the representational capacity of a sample set of limited size by using the lines passing through each pair of the samples belonging to the same class. Simple yet effective, NFL shows good performance in many applications, includ- ing face recognition [1] [2], audio retrieval [3], image classification [4], speaker identification [5] and object recognition [6]. On the other hand, feature lines may produce detrimental effects that lead to increased decision errors. Compared with the well-known Nearest Neighbor (NN) classifier [7], NFL has obvious drawbacks under certain situations that limit its further potential. The authors of [8] pointed out one of the problems – extrapolation inaccuracy, and proposed a solution called Nearest Neighbor Line (NNL). This extrapolation inaccuracy may lead to enormous decision errors in a low dimensional feature space while a simple NN classifier easily reaches a perfect correct classification rate of 100%. Another drawback of NFL is interpo- lation inaccuracy. Distributions assuming a complex shape (two-spiral problem for example) often fall into this category, where, by the original NFL, the inter- polating parts of the feature lines of one class break up the area of another class and severely damage the decision region. ⋆ Corresponding author. Email: chenyq@fudan.edu.cn