Debiasing weighted multi-view k-means clustering based on causal regularization

Abstract

In the field of unsupervised learning, many methods such as clustering rely on exploring the correlations among features. However, considering these correlations is not always advantageous for learning models. The biased selection of data may lead to redundant and unstable correlations among features, adversely affecting the performance of learning models. Multi-view data presents more complex feature correlations with potential redundancy and varying distributions across views, necessitating detailed analysis. This paper proposes a causal regularized debiased multi-view k-means clustering (DMKC) method to counteract redundant feature correlations stemming from sample selection bias. This method introduces a covariate weighted balance method from causal inference to mitigate redundant bias in multi-view clustering by adjusting sample weights. The approach combines sample and view weights within a k-means loss framework, effectively eliminating feature redundancy and enhancing clustering performance amidst sample selection bias. The optimization process of the relevant parameters is detailed in this paper, and comprehensive experiments demonstrate the effectiveness of the method.