Incomplete Multi-View Learning Under Label Shift.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Incomplete multi-view learning faces challenges due to data uncertainty and diversity.
• Existing methods often overlook label shift, where training and testing label distributions differ.
• This divergence complicates annotation and model generalization.

Purpose of the Study:

• To propose a novel framework, Incomplete Multi-view Learning under Label Shift (IMLLS), for handling incomplete multi-view data with label shift.
• To formally define IMLLS and a bidirectional complete representation capturing intrinsic and common structures.
• To develop a method that aligns label distributions and improves classification performance.

Main Methods:

• A multilayer perceptron combining reconstruction and classification loss learns latent representations.
• Theoretical proofs establish the existence, consistency, and universality of the learned representation under label shift.
• A novel estimation scheme calculates importance weights to align label distributions, balancing finite sample errors.
• Fine-tuning the reweighted classifier reduces the gap between source and target representations.

Main Results:

• The proposed IMLLS framework demonstrates superior performance compared to state-of-the-art methods.
• The algorithm effectively handles incomplete multi-view data and addresses the label shift problem.
• Experimental validation includes successful discrimination between schizophrenic patients and healthy controls.

Conclusions:

• The IMLLS framework provides an effective solution for incomplete multi-view learning scenarios with label shift.
• The method offers theoretical guarantees and practical improvements in classification tasks.
• The approach shows promise for real-world applications, including medical diagnoses.