Robust resonant anomaly detection with NPLM.

Area of Science:

• Particle Physics
• Machine Learning
• Anomaly Detection

Background:

• Standard methods like Boosted Decision Trees (BDTs) with CWoLa face challenges in detecting rare signal events.
• Existing approaches often require prior assumptions about signal models, limiting their applicability.
• Accurate background modeling is crucial but not always feasible in experimental setups.

Purpose of the Study:

• To evaluate the New Physics Learning Machine (NPLM) algorithm as an alternative to BDTs for anomaly detection.
• To explore NPLM's effectiveness in scenarios with rare signals and unreliable background models.
• To assess NPLM's potential for improving detection performance and reducing uncertainty in particle physics.

Main Methods:

• Investigated NPLM's end-to-end application for anomaly detection and hypothesis testing.
• Utilized in-sample evaluation of a binary classifier to estimate log-density ratios.
• Examined two NPLM approaches: direct application with reliable background and classifier use with hyper-testing for threshold optimization.

Main Results:

• NPLM-based methods demonstrated superior detection performance over BDT-based approaches, particularly in low signal scenarios.
• NPLM significantly reduced epistemic variance associated with hyperparameter choices.
• The NPLM classifier approach, combined with hyper-testing, enhanced performance when background modeling was uncertain.

Conclusions:

• NPLM presents a promising alternative for robust resonant anomaly detection in particle physics.
• The algorithm improves sensitivity and consistency, even with signal variability and uncertain background models.
• This study lays the groundwork for future NPLM-based methods in high-energy physics research.