ESIMCE: Efficient and simple incomplete multi-view clustering via ensembles

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Previous incomplete multi-view clustering (IMVC) methods face challenges with computational complexity, hyper-parameter tuning, missing data imputation, and early information fusion.
• These limitations hinder the scalability and effectiveness of existing IMVC techniques on large, complex datasets.

Purpose of the Study:

• To propose an efficient and robust IMVC method, termed Efficient and Simple Incomplete Multi-view Clustering via Ensembles (ESIMCE).
• To address the limitations of high computational complexity, intractable hyper-parameter tuning, poor imputation of missing information, and suboptimal early-stage information fusion in existing IMVC methods.

Main Methods:

• ESIMCE constructs partial bipartite (anchor) graphs for incomplete views using a shared anchor set.
• It recovers missing data by leveraging cross-view complementary information and sparsifies graphs via K-nearest anchors.
• The method fuses multiple base clusterings at the partition-level using a unified bipartite graph for efficient final partitioning.

Main Results:

• ESIMCE achieves near-linear time complexity, making it suitable for large-scale problems.
• The method demonstrates effective imputation of missing information by exploiting cross-view consistency.
• Experiments show ESIMCE's robustness and efficiency on real-world multi-view datasets, outperforming existing methods.

Conclusions:

• ESIMCE offers an efficient and robust solution for incomplete multi-view clustering.
• The proposed partition-level fusion strategy effectively overcomes the drawbacks of early-stage fusion.
• ESIMCE provides a scalable and effective approach for handling incomplete multi-view data without dataset-specific parameter tuning.