关键词: 子空间聚类, 离群点, 自动权重, 块对角方法

Abstract: Aiming at the problem  that traditional spectral clustering-based subspace clustering methods were prone to outlier interference and thus show degraded clustering performance when there were outliers in high-dimensional data, we proposed an active block diagonal subspace clustering method based on automatic weighting. The method first assigned a corresponding weight to each data point, identified outliers in the data through weight differences, then actively reduced its contribution in the representation matrix to construct a better representation matrix and improved the  clustering performance of the model. Experimental results on 10 datasets compared with 8  algorithms show that the average clustering accuracy, normalized mutual information, and adjusted Rand index of the proposed method are generally better than the comparison algorithms on datasets with 10% or 20% outliers. It performs the best or ranks in the top three on more than half of the datasets in general clustering tasks. Therefore,  the method can not only efficiently handle high-dimensional data clustering with outliers, but also maintain competitiveness in general clustering tasks, providing an effective solution to enhance the robustness of high-dimensional data clustering and having high practical application value.

Key words: subspace clustering, outliers, automatic weighting, block diagonal method