Combining outlier analysis algorithms to identify new physics at the LHC

TL;DR

This paper compares various anomaly detection algorithms, including autoencoders and traditional models, applied to simulated LHC data, and finds that combining scores in the latent space enhances new physics discovery potential.

Contribution

It introduces a method of combining anomaly scores from multiple algorithms, especially in the VAE latent space, to improve detection of new physics signals at the LHC.

Findings

VAE-based algorithms outperform others in anomaly detection.

Combining scores in the latent space improves detection accuracy.

Logical AND of scores from VAE latent space algorithms is most effective.

Abstract

The lack of evidence for new physics at the Large Hadron Collider so far has prompted the development of model-independent search techniques. In this study, we compare the anomaly scores of a variety of anomaly detection techniques: an isolation forest, a Gaussian mixture model, a static autoencoder, and a $\beta$-variational autoencoder (VAE), where we define the reconstruction loss of the latter as a weighted combination of regression and classification terms. We apply these algorithms to the 4-vectors of simulated LHC data, but also investigate the performance when the non-VAE algorithms are applied to the latent space variables created by the VAE. In addition, we assess the performance when the anomaly scores of these algorithms are combined in various ways. Using supersymmetric benchmark points, we find that the logical AND combination of the anomaly scores yielded from algorithms trained in the latent space of the VAE is the most effective discriminator of all methods tested.