Fuzzy One-Class Extreme Auto-encoder

摘要

A novel one class classification (OCC) algorithm called fuzzy one class extreme auto-encoder (FOCEAE) is presented in this article. The algorithm combines the precision of probability density estimation and the generalization of neural networks to accurately generate the compact bound for the target class cases. Firstly, a K-nearest-neighbors non-parametric probability density estimation-alike strategy is used to estimate the relative densities of all target class training objects, then the relative densities are transformed to be the fuzzy coefficients for further training fuzzy extreme learning machine (FELM) model. Specifically, considering there are only one-class instances, FELM is trained in the form of auto-encoder, i.e., each input equals to be the expected output of the network. Finally, the bound (i.e., the threshold) of the target class cases is determined by calculating and ranking the reconstructed errors of all training instances. We show the effectiveness and superiority of the proposed FOCEAE algorithm by comparing it with some benchmark OCC algorithms on a mass of data sets in terms of both F-measure and G-mean metrics. The statistical results also indicate that the proposed algorithm performs significantly better than some conventional ones.