An Efficient Method for Outlying Aspect Mining Based on Genetic Algorithm

Zihao Chen, Lei Duan, Xinye Wang

Abstract: Outlying aspect mining (OAM) aims to identify a feature subspace in which a given query object is dramatically distinctive from the rest data. The identified features can assist the formulation and optimization of decisions. Score-and-search methods are widely used in outlying aspect mining. However, limited by scoring instability and search inefficiency, studies using this strategy are unable to be comprehensive and accurate for mining outlying aspects. In this paper, it proposes a novel OAM method based on genetic algorithm, named OSIER, which can be applied in mining outlying aspects from multi-dimensional spaces. OSIER improves the search efficiency by analyzing the correlations between dimensions. By combining the genetic algorithm with the traditional beam search strategy, OSIER effectively improves the diversity of the searched aspects. As a result, the execution time for candidate outlying aspects search is controlled in an acceptable range. Experiments show that OSIER outperforms the benchmark methods in terms of effectiveness on the OAM task. Besides, OSIER is capable of providing valuable outlying aspect mining results for various types of datasets.