A New Method to Calculate Weights of Attributes in Spectral Clustering Algorithms

In order to improve the spectral clustering algorithms, there is one algorithm that when calculate the distance formula as similar matrix of spectral clustering, considering the weights of different attributes, but the methods to calculate weights of attributes are of significant limitation, affecting the clustering results. Consequently, inspired by the MIV method, we construct a method named MDIV that is to calculate the weights of attributes by the average distance change from attribute values change. For illustration, two UCI data sets are utilized to show the feasibility of the MDIV in calculating weights of attributes of spectral clustering algorithms. Empirical results show that attributes weights determined by MDIV improve the spectral clustering algorithm more significantly than by factor analysis. That is to say, weights of attributes calculated by the MDIV method are more realistic than other common methods. The MDIV method can effectively reduce human intervention in calculating weights of attributes, and obtain more objective results to improve accuracy of clustering. So, to a certain extent, MDIV can be a general method to calculate weights of attributes in clustering.