A New Method for Components Similarity Measure in Gaussian Spectral Clustering

Considering that the adjacent matrix estimated via the Bhattacharyya distance does not match well the distributions of a Gaussian mixture model, which leads to lower accuracy in Gaussian spectral clustering for the imbalanced and large-scale data. Based on the Bhattacharyya distance combined with Matusia’s measure, a novel estimated components overlap coefficient, the Bhattacharyya distance weighted by the geometric mean of the mixture coefficients, is proposed to enhance the clustering ability in Gaussian mixture model. Compared to Bhattacharyya distance, the efficiency of this method is evaluated via not only the contour plot of the GMM but accuracy using GMM spectral clustering for the imbalanced and large-scale data. Experimental results show that it outperforms traditional methods, yielding more accurate clustering outcomes.