Reproduce “Simple Yet Effective Selective Imputation for Incomplete Multi-view Clustering” on OpenResearch

About this paper

Cai Xu, Jinlong Liu, Yilin Zhang, Ziyu Guan, Wei Zhao, Xiaofei He

SI^3, a method for incomplete multi-view clustering, quantifies the "imputation-relevant informativeness" of missing data points before selectively imputing them. This approach integrates uncertainty-aware imputation at the latent distribution level within a variational inference framework, achieving superior clustering performance, particularly with high or unbalanced missing rates, and demonstrates a substantial computational speedup over existing cautious imputation methods.

Key insights

• Evaluating the "imputation-relevant informativeness" *before* imputation reduces noise introduction and improves computational efficiency compared to post-imputation assessment strategies.
• Imputing missing values as distributions, rather than single point estimates, allows for explicit modeling of uncertainty, which enhances the robustness of multi-view fusion and learned latent representations.
• Combining both intra-view (local neighborhood) and cross-view (semantic consistency) evidence provides a comprehensive and reliable basis for assessing the potential quality of an imputation.

Summary by alphaXiv.