PASCL: Supervised Contrastive Learning with Perturbative Augmentation for Particle Decay Reconstruction

TL;DR

This work uses a compact matrix representation termed as lowest common ancestor generations matrix, to encode the particle decay tree structure, and introduces a perturbative augmentation technique applied to node features, aiming to mimic experimental uncertainties and increase data diversity.

Abstract

In high-energy physics, particles produced in collision events decay in a format of a hierarchical tree structure, where only the final decay products can be observed using detectors. However, the large combinatorial space of possible tree structures makes it challenging to recover the actual decay process given a set of final particles. To better analyse the hierarchical tree structure, we propose a graph-based deep learning model to infer the tree structure to reconstruct collision events. In particular, we use a compact matrix representation termed as lowest common ancestor generations (LCAG) matrix, to encode the particle decay tree structure. Then, we introduce a perturbative augmentation technique applied to node features, aiming to mimic experimental uncertainties and increase data diversity. We further propose a supervised graph contrastive learning algorithm to utilize the information of inter-particle relations from multiple decay processes. Extensive experiments show that our proposed supervised graph contrastive learning with perturbative augmentation (PASCL) method outperforms state-of-the-art baseline models on an existing physics-based dataset, significantly improving the reconstruction accuracy. This method provides a more effective training strategy for models with the same parameters and makes way for more accurate and efficient high-energy particle physics data analysis.

Figures (4)

• Figure 1: A simulated particle decay event from the Phasespace dataset with 6 stable particles (a modified version of figure 2 in LCAG). The tree structure shows an example of particle decay process (left), and the matrix represents its corresponding LCAG matrix (right). Labels and colors are only used as a symbol for discrimination.
• Figure 2: $\textbf{Overview of PASCL{}}$. $\textbf{a}$: Node features perturbative augmentation. We use three perturbative augmentation methods: masking, adversarial perturbation and Gaussian perturbation to construct the augmented particle graph $\widetilde{\mathcal{G}}$. $\textbf{b}$: Relations construction and embedding. Extract the representations of $h$ and $\widetilde{h}$ between nodes in original particle graph $\mathcal{G}$ and the augmented particle graph $\widetilde{\mathcal{G}}$, using MLP-based message passing and aggregation. $\textbf{c}$: Supervised contrastive learning on graph. We establish the connection between different decay processes by constructing positive and negative sample pairs among them. The graph encoder and projection head is then trained to maximize the consistency between the same relations.
• Figure 3: Performance of PASCL with different perturbative augmentation methods. We vary the degree of different perturbations in data augmentation and report the corresponding performance trends across different parameters, represented by different colors. The red represents the best performing model, blue represents the worst, and the gray dashed line represents the baseline. The horizontal axis represents the number of leaves up to and including, and the vertical axis represents the accuracy in the classification task (higher is better). Results are expressed as the average of ten separate runs.
• Figure 4: Alignment analysis. We show t-SNE visualizations of inter-particle relationship representations during 50 random decay processes to investigate the correlation of splitting at the same level. The colors represent the level of splitting (i.e. the value of the edge in the truth LCAG matrix). From top to bottom are the cases of 8, 12, and 16 leaf nodes, respectively.