Topology-Preserved Information Bottleneck for Multiview Anomaly Detection

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Anomaly detection (AD) is crucial across various domains.
• Existing AD methods often rely on single-view data, limiting their effectiveness with complex multiview datasets.
• Multiview data offers richer information but poses challenges for traditional AD techniques due to feature overlap and fusion complexities.

Purpose of the Study:

• To develop an effective anomaly detection method for multiview data.
• To address limitations of existing methods in handling feature overlaps and preserving data structure during fusion.
• To propose a novel feature extraction technique that is inherently capable of anomaly detection.

Main Methods:

• Leveraging the information bottleneck (IB) principle to extract concise and comprehensive representations from multiview data.
• Designing a topology-preserved regularization to maintain the intrinsic structure of the original data in latent representations.
• Developing the topology-preserved multiview information bottleneck (TMVIB) feature extraction method.

Main Results:

• The proposed TMVIB method effectively extracts concise, comprehensive, and topology-preserving latent representations.
• The TMVIB feature extraction method demonstrates inherent anomaly detection capabilities, directly outputting anomaly scores.
• Experiments on synthetic and real-world multiview datasets validate the effectiveness of the TMVIB approach.

Conclusions:

• The TMVIB method offers a robust solution for anomaly detection in multiview data.
• Preserving data topology during feature extraction is critical for accurate anomaly detection.
• The TMVIB framework provides a unified approach for feature extraction and anomaly scoring in multiview settings.