Incorporating Physical Priors into Weakly Supervised Anomaly Detection.

Area of Science:

• Machine Learning
• High Energy Physics
• Data Analysis

Background:

• Traditional anomaly detection struggles with rare signals and high-dimensional, noisy data.
• Weakly supervised methods lack sensitivity when signal models are not precisely known.
• Existing approaches degrade performance significantly with irrelevant input features.

Purpose of the Study:

• To introduce a novel machine learning strategy for anomaly detection using weak supervision.
• To enhance search sensitivity in scenarios with rare signals or numerous unhelpful features.
• To develop a method robust against irrelevant input dimensions (noise).

Main Methods:

• Proposed prior-assisted weak supervision (PAWS), a machine learning-based anomaly detection strategy.
• Incorporated information from a class of signal models into the weak supervision framework.
• Utilized a mix of semisupervised and weakly supervised learning techniques.

Main Results:

• PAWS significantly enhances the search sensitivity of weakly supervised anomaly detection.
• Achieved a factor of 10 increase in sensitivity (cross-section) on the LHC Olympics dataset compared to previous methods.
• Demonstrated robustness to irrelevant input dimensions, maintaining performance where classical methods degrade by another factor of 10.

Conclusions:

• PAWS matches the sensitivity of fully supervised methods without requiring exact parameter specification.
• The method pushes the frontier of sensitivity, bridging model-agnostic and model-specific anomaly searches.
• PAWS offers a powerful new approach applicable to various anomaly detection scenarios.