PGFormer: A Prototype-Graph Transformer for Incomplete Multiview Clustering.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Incomplete multiview clustering (IMVC) is challenged by missing data and view discrepancies.
• Existing deep learning methods for IMVC often create biased representations and inaccurate imputations by forcing view representations to be identical.

Purpose of the Study:

• To propose a novel IMVC framework, the prototype-graph transformer (PGFormer), that enhances clustering performance by integrating prototype assignments.
• To address limitations of existing methods by improving representation learning and data imputation in IMVC.

Main Methods:

• PGFormer utilizes view-specific encoders and a graph convolutional network (GCN) to model topologies and generate prototypes.
• Dual attention mechanisms (prototype-to-prototype self-attention and prototype-to-node cross-attention) refine embeddings and explore topological relationships.
• A cross-prototype imputation (CPI) module addresses missing data using weighted prototype assignments, and a cross-view alignment module ensures consistent predictions.

Main Results:

• PGFormer demonstrates superior performance compared to existing baseline methods in incomplete multiview clustering tasks.
• The framework effectively refines node embeddings and reconstructs available samples using these refined embeddings.
• The proposed methods successfully address biased representations and inaccurate imputation issues.

Conclusions:

• The prototype-graph transformer (PGFormer) offers a significant advancement in incomplete multiview clustering.
• PGFormer's novel approach of integrating prototype assignments effectively handles missing data and view discrepancies.
• The framework shows promise for improving the accuracy and robustness of clustering algorithms in complex, incomplete datasets.