Anomalous jet identification via sequence modeling

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