Scalable incomplete multi-view clustering with adaptive data completion

Existing multi-view clustering algorithms are typically built on the prior assumption that data collected from different sources are complete without missing, while this is not always satisfied in real-world applications. Incomplete multi-view clustering (IMC) aims at discovering the latent cluster structure and partitioning the incomplete multi-view data into different groups, which is more practical yet challenging. Moreover, the main bottleneck of the current IMC research is how we could economically cluster large-scale incomplete multi-view data with limited resources. In this paper, we propose a novel Scalable Incomplete Multi-view Clustering with Adaptive Data Completion (SIMC_ADC) method, which seeks to deal with large-scale IMC with promising instance-level restoration. Specifically, to cater to large-scale IMC, our SIMC_ADC unifies representative anchor learning and similarity recovery into a one-stop learning scheme within linear computational and memory costs. Moreover, we formulate an adaptive instance completion scheme to discover and generate trustworthy underlying connections across instances, which iteratively infers and pads the missing instances along with the complete similarity learning. Importantly, theoretical analysis on data completion error further guarantees the reliability of such data completion paradigm. Extensive experiments validate the efficiency and effectiveness of our method when handling large-scale IMC problems in comparison with state-of-the-art algorithms. The source code for our SIMC_ADC is available at https://github.com/DarrenZZhang/SIMC_ADC.