One-class classifier based on principal curves

Abstract

One-class classification is a special multi-class approach where data from only a single class are available for classifier training. It is an approach with several applications in real-world scenarios, for instance, for outlier or novelty detection. This paper presents a new one-class classifier based on principal curves. The method exploits the good capacity of data representation of the principal curves to build a compact data representation of the known class. The use of principal curves gives the proposed method a good capacity for dealing with different shapes of the feature space, leading to better performance rates. The results showed high performances of the proposed method for synthetic and real data sets, outperforming other known one-class learning algorithms. Moreover, it builds decision boundaries more uniform around the known class than other models and is a fast method during the operating stage since classification is performed by simply mapping the Euclidean distances from data to the principal curve.