Exploring Optimal Transport for Event-Level Anomaly Detection at the Large Hadron Collider
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.
Kai: We're looking at the paper: "Exploring Optimal Transport for Event-Level Anomaly Detection at the Large Hadron Collider".
Mira: Here's the gist of the paper. Anomaly detection is explored using balanced Optimal Transport distances to find physics beyond the Standard Model signatures at the Large Hadron Collider The choice of ground space plays a key role in optimizing performance The work investigates using balanced OT for anomaly detection on L1 trigger-level data from the CMS anomaly detection challenge dataset (CMS ADC 2021) Optimal Transport Framework Optimal Transport theory provides a framework for comparing probability distributions by assigning a cost to transform one probability distribution into another The cost of moving a differential piece of the source distribution to construct the target distribution is the amount of probability mass weighted by (an integer power of) how far that differential piece has to move in the ground space The collection of these differential movements is called a transport plan, and the net cost of executing the optimal (lowest-cost) transport plan induces a distance between the source and target distributions A mathematically robust choice of ground space cost metric is the squared Euclidean distance, c(x, x′) = x − x′2 The distance associated with the solution of the OT problem with this cost metric is the 2-Wasserstein distance The 2-Wasserstein distance is given by W2 (E, E′) = min g∈Γ (E,E′) X i,j gi jxi − x'j2!1/2 Anomaly Scoring and Ground Space The key expectation is that the 2-Wasserstein distance between two SM events should generally be less than the distance between a SM event and a BSM event The anomaly score for test event T is chosen to be the minimum of N 2-Wasserstein distance values, i.e., min of N W2 (T,S1), each S1 being an SM reference event The choice of a ground space that maximizes the differences between SM and BSM events (and minimizes differences between SM events) is essential Past works have opted for a physically-motivated ground space where the ground space is the relative η,φ plane centered on a jet’s 4-vector location, which is called the “2D” ground space In this work, an alternate “3D” ground space which adds an additional pT/GeV dimension to the standard η,φ plane is considered This choice of a less-intuitive ground space indirectly restores information about the total pT in an event, which is crucial for detecting anomalies Machine Learning Integration The methods considered range from unsupervised to weakly supervised approaches The k-Nearest Neighbors (kNN) algorithm provides a simple and interpretable supervised machine learning strategy for binary classification The one-class Support Vector Machines (SVMs) are fully unsupervised and operate by drawing a decision boundary around the training SM data, implicitly learning to estimate the density of SM events For kNNs, anomaly augmented SM events are used as “signal” during training to promote them to weakly supervised anomaly detection Performance Comparison The performance comparison across different methods shows that the 3D OT distance outperforms the 2D OT distance alone The inclusion of constituent pT information in the 3D OT distance leads to dramatically improved performance in anomaly detection, particularly for the h± signal The 3D OT+anomaly kNN method is comparable to the direct use of 3D OT distances for the A signal model However, fully unsupervised one-class SVM is less competitive than either the 3D OT distance or 3D OT+anomaly kNN methods Feasibility and Future Work In principle, OT methods may be used for anomaly detection both at trigger level and in offline analyses The main considerations governing the feasibility of using OT methods in an online setting are memory limitations and computational evaluation time Calculating the OT distance between two events is computationally costly, but many approximation strategies can lower this computational burden Using linearized OT can reduce the computational cost by computing a single 3D OT distance to a suitably-chosen reference event Future work should incorporate more information into the optimization problem, such as adding additional ground space dimensions like multiplicity or encoding species-type information Another strategy is to improve the correspondence between physical differences and OT distances by accounting for symmetry transformations of the underlying ground space The Gromov-Wasserstein distance is mentioned as a way to handle symmetry transformations, as it only depends on the relative locations of the probability masses within each event The 2-Wasserstein distance based on a three-dimensional ground space that includes pT /GeV outperforms the 2-Wasserstein distance using the η,φ coordinates alone The simple OT distance-based approach demonstrates the key features and potential advantages of using optimal transport for anomaly detection in collider data The paper concludes that anomaly detection based directly on the 3D OT distance outperforms methods that couple it to simple machine learning algorithms such as kNN or one-class SVM However, this approach does not fully match the performance of more sophisticated ML methods on the same dataset The work suggests that there is additional information in signal and background events that is not captured by the simple OT-based methods presented here Appendix Summary The appendix summarizes baseline performance using Total pT, MET, and Total Multiplicity for comparison against recent autoencoder methods The comparison with classification performance shows that the 3D ground space contains much more information than its 2D counterpart The results indicate that for some signal models, there can be extra discriminating information contained in the η,φ distribution of an event For all signal cases, the 3D OT+kNN classification performance outperforms the 3D OT+kNN anomaly detection performance The maximum SI value is found in this signal efficiency range (i.e.
Kai: εs ∈ [0.
Title and authors: Kai: Before we get into the details, let's just talk about who wrote this. The authors are Nathaniel Craig, Jessica N Howard, and Hancheng Li from the Department of Physics at UC Santa Barbara and Kavli Institute for Theoretical Physics there.
Mira: They bring a strong background in theoretical physics and condensed matter concepts to this problem because they are focusing on the assumptions behind the model-agnostic search strategy.
Lev: From an engineering standpoint, I wonder how much computational overhead this optimal transport approach actually takes when you try to run it on real hardware for trigger level data.
Kai: The paper is interested in commenting on the feasibility of implementing these methods in systems like the L1 trigger system, which is a pretty crucial practical question for any new analysis technique.
Mira: They are also looking at how the choice of ground space directly impacts performance, which suggests that just picking a distance function isn't enough; you have to pick the right geometric setup.
Lev: I’m curious about their suggestion regarding that ground space—if they find a way to make it work faster on a real system, that would be a big win for us in terms of overhead.
The paper's summary: Kai: So, the main takeaway is that the goal is to use this OT distance as an anomaly score for event-level detection, and they’re comparing different ground spaces to see which one works best.
Mira: They specifically expect that the two-Wasserstein distance between two Standard Model events will be smaller than the distance between a Standard Model event and some kind of Beyond the Standard Model event <ref:2401.15542#pg1>.
Lev: That makes sense if you think about what an anomaly score is supposed to do; it should be higher for something weird and lower for something normal.
Kai: The paper suggests that a physically motivated ground space, like the relative eta phi plane centered on a jet’s four-vector location, might be good, but they test an alternative three-dimensional ground space by adding an extra pT over GeV dimension.
Mira: Adding that transverse momentum information is what they argue indirectly restores important information about the total pT of an event, which is really crucial for detecting these anomalies.
The paper's improvements: Kai: They also discuss a few ways to improve the approach. One idea mentioned is using k-Nearest Neighbors, or kNNs, with OT distances as an input for binary classification tasks.
Lev: KNN sounds simple and interpretable for classification, but I wonder how you train that system when your training data is already being augmented with anomalies.
Mira: The paper suggests that if you use anomaly augmented Standard Model events as the signal during training, it can help promote those specific anomalies to become better signals for the anomaly detection method.
Kai: They also look at other ground space options, suggesting a more complex three-dimensional setup is better than just using the two dimensional eta phi coordinates alone.
Lev: If you move to that three-dimensional space, does the computational cost jump significantly? We need to know if that complexity translates into a practical speedup for trigger systems.
Conclusion: Mira: To wrap things up, they show that the three dimensional OT distance outperforms just using the two dimensional eta phi coordinates by a large margin.
Kai: The paper concludes that this direct use of the three dimensional OT distance is better than coupling it with simple machine learning algorithms like kNN or one class SVMs for anomaly detection on this dataset.
Lev: That’s an important distinction because it means we don't necessarily need to couple the OT distance with a complex machine learning model to get the best results.
Mira: They also suggest that there might be additional information in both signal and background events that these simple OT-based methods don't capture.
Kai: So, in summary, the paper on "Exploring Optimal Transport for Event-Level Anomaly Detection at the Large Hadron Collider" shows that using a three dimensional ground space including pT over GeV is better than just eta phi coordinates alone.
Lev: It’s a solid finding, but I still think we need to figure out how to make this kind of distance calculation fast enough for real-time trigger decisions.
Mira: Exactly. And that opens up the next question: can we really scale this to even more complex data or different physical signatures?
Nathaniel Craig, Jessica N. Howard, Hancheng Li
Department of Physics, University of California, Santa Barbara · Kavli Institute for Theoretical Physics
hep-ph
Submitted: 2024-01-28
Updated: 2026-11-12
Comments: 26 pages, 4 tables, 7 figures; added additional studies and fixed several typographical errors
Code: https://github.com/hanchengli/anomaly_detection_code
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 76/100
The gist: Anomaly detection is explored using balanced Optimal Transport distances to find physics beyond the Standard Model signatures at the Large Hadron Collider The choice of ground space plays a key role…
Key concepts
- Optimal Transport (OT) Framework
- OT measures the minimum cost to transform one probability distribution of events into another. It assigns a cost based on how far each small piece of probability mass must move in a chosen ground space. The resulting minimum total movement defines the distance between the two event distributions.
- 2-Wasserstein Distance ($W_2$)
- This is the specific distance metric derived from OT using squared Euclidean distance as the cost. It quantifies how different two probability distributions are by finding the most efficient way to transport one set of events onto another, minimizing the total weighted squared distance traveled.
- Ground Space
- The ground space is a mathematical coordinate system used to define where events are located for comparison. The choice of this space—like the standard $\eta,\phi$ plane or a 3D version including $p_T/GeV$—is critical because it dictates what physical differences the OT distance will be sensitive to.
- Anomaly Scoring
- An anomaly score is calculated by finding the minimum 2-Wasserstein distance between a test event and several known Standard Model (SM) reference events. A larger distance suggests the test event is more likely to be a BSM signal, as SM events are expected to be close together in this metric.
Terminology
Summary
Anomaly detection is explored using balanced Optimal Transport distances to find physics beyond the Standard Model signatures at the Large Hadron Collider The choice of ground space plays a key role in optimizing performance The work investigates using balanced OT for anomaly detection on L1 trigger-level data from the CMS anomaly detection challenge dataset (CMS ADC 2021)
Optimal Transport Framework
Optimal Transport theory provides a framework for comparing probability distributions by assigning a cost to transform one probability distribution into another The cost of moving a differential piece of the source distribution to construct the target distribution is the amount of probability mass weighted by (an integer power of) how far that differential piece has to move in the ground space The collection of these differential movements is called a transport plan, and the net cost of executing the optimal (lowest-cost) transport plan induces a distance between the source and target distributions A mathematically robust choice of ground space cost metric is the squared Euclidean distance, c(x, x′) = x − x′2 The distance associated with the solution of the OT problem with this cost metric is the 2-Wasserstein distance The 2-Wasserstein distance is given by W2 (E, E′) = min g∈Γ (E,E′) X i,j gi jxi − x'j2!1/2
Anomaly Scoring and Ground Space
The key expectation is that the 2-Wasserstein distance between two SM events should generally be less than the distance between a SM event and a BSM event The anomaly score for test event T is chosen to be the minimum of N 2-Wasserstein distance values, i.e., min of N W2 (T,S1), each S1 being an SM reference event The choice of a ground space that maximizes the differences between SM and BSM events (and minimizes differences between SM events) is essential Past works have opted for a physically-motivated ground space where the ground space is the relative η,φ plane centered on a jet’s 4-vector location, which is called the “2D” ground space In this work, an alternate “3D” ground space which adds an additional pT/GeV dimension to the standard η,φ plane is considered This choice of a less-intuitive ground space indirectly restores information about the total pT in an event, which is crucial for detecting anomalies
Machine Learning Integration
The methods considered range from unsupervised to weakly supervised approaches The k-Nearest Neighbors (kNN) algorithm provides a simple and interpretable supervised machine learning strategy for binary classification The one-class Support Vector Machines (SVMs) are fully unsupervised and operate by drawing a decision boundary around the training SM data, implicitly learning to estimate the density of SM events For kNNs, anomaly augmented SM events are used as “signal” during training to promote them to weakly supervised anomaly detection
Performance Comparison
The performance comparison across different methods shows that the 3D OT distance outperforms the 2D OT distance alone The inclusion of constituent pT information in the 3D OT distance leads to dramatically improved performance in anomaly detection, particularly for the h± signal The 3D OT+anomaly kNN method is comparable to the direct use of 3D OT distances for the A signal model However, fully unsupervised one-class SVM is less competitive than either the 3D OT distance or 3D OT+anomaly kNN methods
Feasibility and Future Work
In principle, OT methods may be used for anomaly detection both at trigger level and in offline analyses The main considerations governing the feasibility of using OT methods in an online setting are memory limitations and computational evaluation time Calculating the OT distance between two events is computationally costly, but many approximation strategies can lower this computational burden Using linearized OT can reduce the computational cost by computing a single 3D OT distance to a suitably-chosen reference event Future work should incorporate more information into the optimization problem, such as adding additional ground space dimensions like multiplicity or encoding species-type information Another strategy is to improve the correspondence between physical differences and OT distances by accounting for symmetry transformations of the underlying ground space The Gromov-Wasserstein distance is mentioned as a way to handle symmetry transformations, as it only depends on the relative locations of the probability masses within each event
Conclusion
The 2-Wasserstein distance based on a three-dimensional ground space that includes pT /GeV outperforms the 2-Wasserstein distance using the η,φ coordinates alone The simple OT distance-based approach demonstrates the key features and potential advantages of using optimal transport for anomaly detection in collider data The paper concludes that anomaly detection based directly on the 3D OT distance outperforms methods that couple it to simple machine learning algorithms such as kNN or one-class SVM However, this approach does not fully match the performance of more sophisticated ML methods on the same dataset The work suggests that there is additional information in signal and background events that is not captured by the simple OT-based methods presented here
Appendix Summary
The appendix summarizes baseline performance using Total pT, MET, and Total Multiplicity for comparison against recent autoencoder methods The comparison with classification performance shows that the 3D ground space contains much more information than its 2D counterpart The results indicate that for some signal models, there can be extra discriminating information contained in the η,φ distribution of an event For all signal cases, the 3D OT+kNN classification performance outperforms the 3D OT+kNN anomaly detection performance The maximum SI value is found in this signal efficiency range (i.e. εs ∈ [0. p. 7
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Ground space optimization for Optimal Transport distances: The system can be improved by systematically selecting ground spaces that
maximize the differences between SM and BSM events (and minimizes differences between SM events)
to derive an anomaly score, as this is essential for performance. -
Incorporating total transverse momentum into the ground space: The AI can be enhanced by using a three-dimensional ground space that includes pT/GeV as an additional dimension alongside eta,phi coordinates, which has been shown to
dramatically improve performance in anomaly detection.
-
Utilizing simple OT distance as the primary score: The system can be improved by adopting
simply using the minimum 3D 2-Wasserstein distance between a test event and a background sample as an anomaly score,
which outperforms methods coupling OT distances to simple machine learning algorithms like kNN or one-class SVM. -
Developing multispecies Optimal Transport: Future work should explore adding ground space dimensions, such as multiplicity or encoding species-type information, to create
multi-species OT
to capitalize on known differences between SM and many BSM signals. -
Accounting for physical symmetries in ground spaces: The system can be improved by implementing
groundspace transformations (e.g. rotations) to the problem as additional optimization steps,
or considering distances like the Gromov-Wasserstein distance, which aremanifestly invariant to translations and rotations of the underlying ground space.
-
Improving training data proxy for signal: For specific models like A, the system can be made more effective by using anomaly augmentations that
capture this property
(total multiplicity) because they make anomaly-augmented backgrounds abetter proxy for the true A signal events.
Sources
- LHC physics dataset for unsupervised New Physics detection at 40 MHz
- The Dark Machines Anomaly Score Challenge: Benchmark Data and Model Independent Event Classification for the Large Hadron Collider
- The LHC Olympics 2020: A Community Challenge for Anomaly Detection in High Energy Physics
- Anomaly detection search for new resonances decaying into a Higgs boson and a generic new particle $X$ in hadronic final states using $\sqrt{s} = 13$ TeV $pp$ collisions with the ATLAS detector
- Dijet resonance search with weak supervision using $\sqrt{s}=13$ TeV $pp$ collisions in the ATLAS detector
- Autoencoders on FPGAs for real-time, unsupervised new physics detection at 40 MHz at the Large Hadron Collider
- Testing a Neural Network for Anomaly Detection in the CMS Global Trigger Test Crate during Run 3
- Spatio-Temporal Anomaly Detection with Graph Networks for Data Quality Monitoring of the Hadron Calorimeter
- Triggerless data acquisition pipeline for Machine Learning based statistical anomaly detection
- Autoencoder-based Anomaly Detection System for Online Data Quality Monitoring of the CMS Electromagnetic Calorimeter
- Autoencoder-based Online Data Quality Monitoring for the CMS Electromagnetic Calorimeter
- Enhancing the hunt for new phenomena in dijet final-states using anomaly detection filters at the High-Luminosity Large Hadron Collider
- Anomaly Detection for Physics Analysis and Less than Supervised Learning
- Machine Learning for Anomaly Detection in Particle Physics
- Anomalies, Representations, and Self-Supervision
- Unravelling physics beyond the standard model with classical and quantum anomaly detection
- QCD or What?
- Searching for New Physics with Deep Autoencoders
- Topological Obstructions to Autoencoding
- Memorizing Normality to Detect Anomaly: Memory-augmented Deep Autoencoder for Unsupervised Anomaly Detection
Related papers
- MadNIS at NLO
- The NANOGrav 15-year data set: A Limit on Continuous Spin Gravitons
- DIS-secting the universe with high-energy neutrinos
- Crowdsourcing Gravitational Waves from Superradiant Axions
- Classification of g-modes for neutron stars with a strong transition: Novel universal relation including slow stable hybrid stars
- Scalar axion field of toroidal electromagnetic pulses