Randomizing greedy ensemble outlier detection with GRASP

Abstract

Ensemble methods have been recently used in many applications of machine learning in different areas. In this context, outlier detection is an area where recently these methods have received increasing attention. This paper deals with randomization in ensemble methods for outlier detection. We have developed a novel algorithm exploiting stochastic local search heuristics to induce diversity in an ensemble outlier detection algorithm. We exploit the capability of the GRASP heuristic to induce diversity into the search process and to maintain a good balance of exploitation and diversification in building the ensemble. The conducted experiments show interesting improvements over the greedy ensemble method and open the path for novel research in this direction.