Anomaly detection in the presence of irrelevant features

Abstract

Experiments at particle colliders are the primary source of insight into physics at microscopic scales. Searches at these facilities often rely on optimization of analyses targeting specific models of new physics. Increasingly, however, data-driven model-agnostic approaches based on machine learning are also being explored. A major challenge is that such methods can be highly sensitive to the presence of many irrelevant features in the data. This paper presents Boosted Decision Tree (BDT)-based techniques to improve anomaly detection in the presence of many irrelevant features. First, a BDT classifier is shown to be more robust than neural networks for the Classification Without Labels approach to finding resonant excesses assuming independence of resonant and non-resonant observables. Next, a tree-based probability density estimator using copula transformations demonstrates significant stability and improved performance over normalizing flows as irrelevant features are added. The results make a compelling case for further development of tree-based algorithms for more robust resonant anomaly detection in high energy physics.

Improving new physics searches with diffusion models for event observables and jet constituents

“…As the classifiers themselves are not the focus on this work, we use the same hyperparameters and architecture used in refs. [17,18] without further optimisation for the high level features, despite promising performance improvements observed with decision trees [96,132] or for mitigating sculpting after applying cuts [20]. This enables easier direct comparison with CurtainsF4F.…”

“…This drop in performance, however, can be expected and is reproducible in the idealised setting. CWoLa classifiers based on neural networks have been observed to lose sensitivity as the dimensionality of the input increases, especially with inputs which are not sensitive to signal processes [20,96,132]. This effect is amplified as the number of signal events in the training data decreases.…”