Consensus Regularized Multi-View Outlier Detection.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multi-view data presents challenges for outlier detection due to complex distributions across views.
• Conventional methods often struggle to generalize beyond two-view settings, increasing complexity and reducing performance.
• Existing approaches typically rely on pairwise constraints, limiting scalability to more than two views.

Purpose of the Study:

• To propose a novel multi-view outlier detection method that overcomes the limitations of existing pairwise constraint approaches.
• To develop a method that is more scalable and performs better when generalizing from two-view to three-view or more data.
• To explicitly characterize different types of outliers using cluster assignments and sample-specific errors.

Main Methods:

• A novel multi-view outlier detection method incorporating consensus regularization on latent representations.
• Explicit characterization of outliers via intrinsic cluster assignment labels and sample-specific errors.
• An optimization solution derived using the augmented Lagrangian multiplier method.

Main Results:

• The proposed method was evaluated on five machine learning datasets, demonstrating effectiveness in multi-view outlier detection.
• The model was successfully tailored to computer vision tasks, including saliency detection and face reconstruction.
• Experimental results confirm the method's effectiveness in both standard outlier detection and extended computer vision applications.

Conclusions:

• The novel consensus regularization approach offers an effective and scalable solution for multi-view outlier detection.
• The method demonstrates strong performance across various datasets and practical computer vision applications.
• This work advances multi-view outlier detection by providing a more generalizable and robust framework.