Late-acceptance and step-counting hill-climbing GP for anomaly detection

Abstract

One-Class Classification (OCC) for anomaly detection is a method for anomaly detection that constructs a classifier from only normal examples. Classifier systems such as Kernel Density Estimation (KDE) and Support Vector Machine (SVM) typically do well at this task, but can be slow when classifying new instances. Previous work has used Genetic Programming (GP) to learn the density from KDE, with results often out-performing those of one-class SVM (OCSVM) and KDE based OCC (OCKDE). However, the search is computationally expensive, and it suffers from a need to tune many parameters. In this paper, we will introduce the Late Acceptance Hill-Climbing (LAHC) and Step Counting Hill-Climbing (SCHC) algorithms as GP alternatives. These are simple hill-climbing algorithms, with specific methods to avoid local optima, and far less parameters to tune. The results demonstrate that the proposed models are competitive with standard GP, and often out-perform OCSVM. Their query-time is much less than that of OCSVM, and does not scale with the size of training data.