An ordered feature recognition method based on ranking separability

Abstract

Ordered and unordered features generally coexist in real-world ordinal classification tasks. For this kind of task, the classification performance can be improved by distinguishing between ordered and unordered features, so it is necessary to recognize the ordered features. But the existing methods recognize the ordered features by measuring the relationship between a single feature and class, ignoring the influence of other features, so cannot accurately mine the sequential relationship. To address this problem, we consider comprehensively the relationships between all features and class, develop a method of recognizing all the ordered features accurately. Firstly a sample grouping method by constructing the virtual feature is designed, and it can help to focus on the features that are difficult to distinguish. Then we design a metric named ranking separability to measure the distance in the order of superiority and inferiority between classes from the perspective of features. And finally the ordered features are identified by judging whether the elements of ranking separability matrix constitute an ordinal arrangement. The experimental analysis shows that the proposed method is effective in recognizing all the ordered features accurately and can improve the classification performance of classifier on mixed data with coexistence of ordered and unordered features.