Trustworthy data recovery for incomplete multi-view learning

Abstract

Incomplete multi-view learning has recently made progress towards more reliable decision-making. Existing methods mainly follow a two-step process: first, conducting data imputation, and then performing opinion aggregation based on evidential deep learning. Although these methods evaluate reliability in the final decision-making phase, they neglect leveraging uncertainty to guide high-quality data imputation. In this paper, we put forward a novel trusted framework termed as Trustworthy Data Recovery for Incomplete Multi-view Learning (TDR-IML) which enables trustworthy data imputation in an uncertain-supervision way. First, we obtain the $k$-nearest neighbor nodes of the incomplete data instance and construct a multivariate Gaussian distribution to model the missing data’s latent space. Then, we perform multiple samplings for the missing data and filter out low-quality samples whose uncertainty exceeds the average uncertainty of all the sampled data. In addition, we refine the opinion decoupling strategy to mitigate semantic ambiguity, thereby improving the extraction of both consistent and complementary opinions. We finally conduct experiments on real-world datasets to validate our model. The code is available on https://github.com/ding6ding/TDR-IMV.