In high-energy physics, machine learning is increasingly applied to event reconstruction, where deterministic clustering algorithms struggle with efficiency and scalability as data rates grow. We develop a Graph Neural Network (GNN) for clustering calorimeter energy deposits in the context of the high luminosity LHC, designed to extend and improve upon deterministic approaches in online reconstruction. An event can contain over 250k hits in the calorimeter detector, corresponding to around 4000 nodes as input for the GNN, and must be reconstructed in real time. A further challenge arises from the imbalance of reconstruction graphs, only about 2% of edges correspond to physically meaningful connections, making robust training and generalization difficult.
\nTo address this, we propose a combined loss function that integrates focal loss for edge classification with contrastive loss applied to the final embedding space. The contrastive component operates on pairs of edges, where node embeddings are perturbed to form positive samples and shuffled to form negative samples. This regularizes the representation, reduces overfitting, and significantly improves clustering efficiency compared to focal loss alone. These results highlight the potential of contrastive learning to advance robust graph-based methods for event reconstruction.
", "identifiers": [], "publication_date": "2025-10-14", "publisher": "CERN", "related_identifiers": [], "resource_type": { "id": "poster", "title": { "cs": "Poster", "de": "Poster", "en": "Poster", "sv": "Poster" } }, "rights": [ { "description": { "en": "The Creative Commons Attribution license allows re-distribution and re-use of a licensed work on the condition that the creator is appropriately credited." }, "icon": "cc-by-icon", "id": "cc-by-4.0", "props": { "scheme": "spdx", "url": "https://creativecommons.org/licenses/by/4.0/legalcode" }, "title": { "en": "Creative Commons Attribution 4.0 International" } } ], "title": "Robust Edge Classification through Contrastive Learning on Graphs for Event Reconstruction" }, "parent": { "access": { "owned_by": { "user": "39926" }, "settings": { "accept_conditions_text": null, "allow_guest_requests": false, "allow_user_requests": false, "secret_link_expiration": 0 } }, "communities": { "default": "da007c73-b7a0-4672-8f99-11faf13bab28", "entries": [ { "access": { "member_policy": "closed", "members_visibility": "public", "record_submission_policy": "open", "review_policy": "closed", "visibility": "public" }, "children": { "allow": false }, "created": "2025-06-18T07:31:01.048123+00:00", "deletion_status": { "is_deleted": false, "status": "P" }, "id": "da007c73-b7a0-4672-8f99-11faf13bab28", "links": {}, "metadata": { "funding": [ { "funder": { "id": "014bj2y47", "name": "Schmidt Family Foundation" } } ], "title": "Next Generation Triggers", "website": "https://nextgentriggers.web.cern.ch/" }, "revision_id": 12, "slug": "nextgentriggers", "updated": "2025-09-29T09:50:50.476550+00:00" } ], "ids": [ "da007c73-b7a0-4672-8f99-11faf13bab28" ] }, "id": "15zsp-53432", "pids": { "doi": { "client": "datacite", "identifier": "10.17181/15zsp-53432", "provider": "datacite" } } }, "pids": { "doi": { "client": "datacite", "identifier": "10.17181/syvpb-qb709", "provider": "datacite" }, "oai": { "identifier": "oai:oai:repository.cern::syvpb-qb709", "provider": "oai" } }, "revision_id": 7, "source_file_key": "MML_GNN_Poster-Final.pdf", "source_file_url": "https://repository.cern/api/records/syvpb-qb709/files/MML_GNN_Poster-Final.pdf/content", "stats": { "all_versions": { "data_volume": 171684247.0, "downloads": 73, "unique_downloads": 55, "unique_views": 55, "views": 64 }, "this_version": { "data_volume": 171684247.0, "downloads": 73, "unique_downloads": 55, "unique_views": 55, "views": 64 } }, "status": "published", "updated": "2025-10-30T13:39:50.028564+00:00", "versions": { "index": 1, "is_latest": true } }