Efficient Feature Selection in the Presence of Outliers and Noises

Although regarded as one of the most successful algorithm to identify predictive features, Relief is quite vulnerable to outliers and noisy features. The recently proposed I-Relief algorithm addresses such deficiencies by using an iterative optimization scheme. Effective as it is, I-Relief is rather time-consuming. This paper presents an efficient alternative that significantly enhances the ability of Relief to handle outliers and strongly redundant noisy features. Our method can achieve comparable performance as I-Relief and has a close-form solution, hence requires much less running time. Results on benchmark information retrieval tasks confirm the effectiveness and efficiency of the proposed method.

• Type: Conference
• Year: 2008
• Where: AIRS
• Authors: Shuang-Hong Yang, Bao-Gang Hu
• Keywords: AIRS 2008, I-Relief Algorithm Addresses, Information Management, Noisy Features, Redundant Noisy Features