Multi-view k-means clustering algorithm with improved initialization strategy Effective multi-view clustering algorithm

This paper proposes an improved initialization strategy for the multi-view k-means clustering algorithm based on the adaptive sparse membership and weight allocation multi-view k-means clustering algorithm (MVASM). The method first performs k-means++ clustering on each view to obtain the labels for each view, and then defines a cross-view consistency objective function. The labels and view weights are optimized iteratively, and the subviews are integrated according to the view weights. The integrated view is then subjected to k-means++ clustering. The similarity between the subviews and the integrated view is calculated using the labels, and the weights are solved using gradient descent. After iteration, the center of the subviews is calculated using the labels of the integrated view. Finally, the initialized labels, centers, and weights are brought into the MVASM algorithm for clustering. Experiments on multiple benchmark datasets show that the algorithm proposed in this paper improves the accuracy of four evaluation metrics (ACC, NMI, Purity, Jaccard) by an average of 18.49%, 37.15%, 6.34%, and 14.87%, respectively.