Anomalous Jet Identification via Sequence Modeling

TL;DR

This paper introduces an unsupervised sequence modeling approach using VRNNs to identify anomalous jets in particle physics, improving robustness and sensitivity in boosted object searches.

Contribution

It presents a novel VRNN-based method for jet anomaly detection that operates without high-level variables and is trained in an unsupervised manner, enhancing robustness against correlations.

Findings

Consistent anomaly detection across various signal contamination levels

Improved signal sensitivity while maintaining background jet mass distribution

Effective in both two- and three-pronged jet substructure scenarios

Abstract

This paper presents a novel method of searching for boosted hadronically decaying objects by treating them as anomalous elements of a contaminated dataset. A Variational Recurrent Neural Network (VRNN) is used to model jets as sequences of constituent four-vectors. After applying a pre-processing method which boosts each jet to the same reference mass and energy, the VRNN provides each jet an Anomaly Score that distinguishes between the structure of signal and background jets. The model is trained in an entirely unsupervised setting and without high level variables, making the score more robust against mass and $p_{T}$ correlations when compared to methods based primarily on jet substructure. Performance is evaluated on the jet level, as well as in an analysis context by searching for a heavy resonance with a final state of two boosted jets. The Anomaly Score shows consistent performance along a wide range of signal contamination amounts, for both two and three-pronged jet substructure hypotheses. Analysis results demonstrate that the use of Anomaly Score as a classifier enhances signal sensitivity while retaining a smoothly falling background jet mass distribution. The model's discriminatory performance resulting from an unsupervised training scenario opens up the possibility to train directly on data without a pre-defined signal hypothesis.