Joint Design of Sample and Node Weighting for Distributed Ordinal Regression Under Label Noise

Ordinal regression (OR) is a category of special classification problem where the labels have natural orders. In some practical OR applications, data may be independently collected by multiple agencies. Due to privacy protection or some other constraints, centralized processing is not feasible, and distributed methods are more suitable. Besides, the collected data may have noisy labels, and the noise levels at each agency could be different because of the difference in data collection environments among agencies. A series of works adopt weighting strategies based on label quality to handle label noise in distributed scenarios. Some of them consider intra-node sample-level weighting while ignoring the difference in label noise levels among nodes. Some others consider node-level weighting while not eliminating the impact of label noise inside a node. In fact, sample weighting and node weighting are interrelated, and can interact with each other to improve overall performance. In this paper, we propose a joint design of sample and node weighting (JSNW) for distributed OR under different levels of label noise across nodes. In JSNW, sample and node weighting interact with each other to provide adaptive sample and node weights for model update to mitigate the impact of label noise. Theoretically, we prove the rank consistency of JSNW. Experimentally, the results demonstrate the effectiveness of the joint design and show that JSNW outperforms several state-of-the-art methods.