Two-Phase Attribute Ordering for Unsupervised Ranking of Multi-attribute Objects

Unsupervised ranking faces a problem of distinguishing those critical attributes to ranking. Prior knowledge of ranking might open a new door for this problem. By embedding the ranking prior information, strictly monotonicity and smoothness, this paper presents a two-phase attribute selection procedure for unsupervised ranking. The first phase identifies those irrelevant attributes based on mean Spearman Ranking Correlation Coefficients (SRCCs) of pairs of attributes by knowing that relevant attributes are assumed to be monotone with each other if it is monotone with the ranking score. The second phase carries out Extended Fourier Amplitude Sensitivity Test (EFAST) on a learned ranking rule and provides the total effect for each attribute to ranking. Finally, the most important attribute to ranking are selected to perform ranking. Numerical experiments on synthetical and real datasets illustrate the effectiveness of the two-phase attribute selection for unsupervised ranking.