Parameter-Free Consensus Embedding Learning for Multiview Graph-Based Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Multiview graph-based clustering aims to find a consensus embedding from multiple data sources.
• A key challenge is managing inconsistencies present across these different views.
• Existing methods often require parameter tuning, limiting practical applicability.

Purpose of the Study:

• To propose a parameter-free model for multiview graph clustering that addresses view inconsistencies.
• To enhance both clustering effectiveness and practical applicability.
• To develop an efficient optimization algorithm for the proposed model.

Main Methods:

• A novel two-layer model is proposed to handle view diversities.
• The first layer addresses feature-level inconsistencies within each view.
• The second layer links pre-embeddings from multiple views attentively using a kernel method.
• An efficient alternative algorithm is developed for optimization.

Main Results:

• The proposed model successfully learns a consensus embedding.
• The learning procedure is entirely parameter-free.
• Experimental evaluations on synthetic and real datasets demonstrate the model's effectiveness.
• The developed algorithm is efficient and practical to implement.

Conclusions:

• The proposed parameter-free model effectively alleviates inconsistencies in multiview graph clustering.
• The two-layer approach captures both feature and view-level diversities.
• The method offers a practical and effective solution for learning consensus embeddings.