Not All, but the Right Ones: Energy-Guided Representation Learning for Incomplete Multiview Clustering.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Incomplete multiview clustering (IMVC) addresses challenges in uncovering shared structures from partially missing data.
• Existing IMVC methods face a trade-off between imputation-free approaches (struggle with high missingness) and imputation-based methods (risk error propagation).

Purpose of the Study:

• To propose a novel selective imputation framework, the energy-guided representation learning network (ERL-Net), for IMVC.
• To adaptively guide feature imputation, fusion, and alignment by leveraging energy-based modeling.

Main Methods:

• ERL-Net employs multiview feature extraction via view-specific autoencoders and a shared projection network.
• Energy-geometric graph encoding evaluates feature reliability using a learnable energy function and models inter-view dependencies.
• Energy-gated imputation selectively reconstructs missing views, retaining only reliable candidates.
• Energy-weighted fusion and alignment integrate observed and imputed features, enforcing semantic consistency.

Main Results:

• ERL-Net demonstrates superior performance in IMVC, particularly under high missing data ratios.
• The framework achieves significant improvements over state-of-the-art methods on multiple benchmark datasets.
• Selective imputation based on energy-based reliability proves effective in handling missing views.

Conclusions:

• ERL-Net offers an effective solution for IMVC by intelligently handling missing data through selective imputation.
• The proposed energy-guided approach mitigates the risks associated with traditional imputation methods.
• ERL-Net advances the field of IMVC, showing robust performance even with substantial data incompleteness.