Feature Selection Method Based on Feature's Classification Bias and Performance

摘要

Feature selection is one of the most important dimension reduction technologies for big data issues. Common feature evaluation criteria measure feature with a global score. There are two shortcomings of this strategy, i.e., partially predominant features submerged and classification redundancies for multi-features estimated inappropriately. In this paper, a new feature selection method based on Classification Bias and Classification Performance is proposed. Classification bias describes feature's inclination for recognizing instances to each class. And classification performance measures the consistency of classification bias with prior classification information for assessing feature's classification ability. Thus feature is reconstructed as classification performance vector. This vectorization representation of feature's classification ability is beneficial for those partially predominant features winning out. Also multi-features' classification redundancies are easily measured in the new space. Experimental results demonstrate the effectiveness of the new method on both low-dimensional and high-dimensional data.