SciPost Submission Page
Rare and Different: Anomaly Scores from a combination of likelihood and out-of-distribution models to detect new physics at the LHC
by Sascha Caron, Luc Hendriks, Rob Verheyen
This Submission thread is now published as
Submission summary
| Authors (as registered SciPost users): | Luc Hendriks · Rob Verheyen |
| Submission information | |
|---|---|
| Preprint Link: | https://arxiv.org/abs/2106.10164v5 (pdf) |
| Date accepted: | 2022-01-13 |
| Date submitted: | 2021-12-24 14:40 |
| Submitted by: | Hendriks, Luc |
| Submitted to: | SciPost Physics |
| Ontological classification | |
|---|---|
| Academic field: | Physics |
| Specialties: |
|
| Approaches: | Computational, Phenomenological |
Abstract
We propose a new method to define anomaly scores and apply this to particle physics collider events. Anomalies can be either rare, meaning that these events are a minority in the normal dataset, or different, meaning they have values that are not inside the dataset. We quantify these two properties using an ensemble of One-Class Deep Support Vector Data Description models, which quantifies differentness, and an autoregressive flow model, which quantifies rareness. These two parameters are then combined into a single anomaly score using different combination algorithms. We train the models using a dataset containing only simulated collisions from the Standard Model of particle physics and test it using various hypothetical signals in four different channels and a secret dataset where the signals are unknown to us. The anomaly detection method described here has been evaluated in a summary paper [1] where it performed very well compared to a large number of other methods. The method is simple to implement and is applicable to other datasets in other fields as well.
Author comments upon resubmission
Thank you for reading our article and for the feedback to our submission. We have updated several paragraphs that incorporates the given feedback. Below our individual responses to the points:
point 1:
This is an excellent remark which is also discussed in the original Deep SVDD paper. The solution proposed there is to remove bias from the neurons and not use 0 as a target output value. However we found that the neural network never converged to this trivial solution. We have added a paragraph to the Deep SVDD section explaining the problem, solution and reasoning.
point 2:
We have added a discussion of this issue to the end of the discussion. In short, mismodelling of the background will indeed lead to worse performance, either through a misclassification of anomalies, or through failing subsequent statistical tests. We want to emphasise that the methods proposed here represent a very minor adjustment to the techniques the experiments currently use, only replacing the conventional signal scores by an anomaly score. As such, any negative effects due to mismodelling of the background would also apply to the current techniques.
point 3:
Background estimation is a powerful way to determine signal regions where we can perform counting experiments as well. However, our method is a way to detect anomalous using an event-by-event scoring system, as was the requirement of the Dark Machines challenge (2105.14027), for which our method is the overall winning one. These methods have the advantage that they allow for the use of the tools that are already widely available and used in the experiments. Anomaly detection that involves background estimation is in some sense an orthogonal approach, and we have added a paragraph in the introduction to clearly distinguish the two.
List of changes
- Added paragraph about dealing with trivial solutions for the Deep SVDD model (page 6)
- Added paragraph about misleading scores due to mismodeling of the data (page 12)
- Added paragraph about background estimation and how those methods relate to ours (page 2)
Published as SciPost Phys. 12, 077 (2022)