Learning to Predict Network Paths: A Transformer Model with Confidence-Based Imputation

¹ University of Michigan Physics, Ann Arbor, MI, USA
² European Organization for Nuclear Research (CERN), Geneva, Switzerland
³ Physics Department, University of Chicago, Chicago, IL, USA
⁴ Faculty of Mathematics and Informatics, University of Plovdiv, Bulgaria

Published online: 7 October 2025

Abstract

The research and education community relies on a robust network to access the vast amounts of data generated by scientific experiments. The underlying infrastructure connects hundreds of sites around the world, requiring reliable and efficient transfers of increasingly large datasets. These activities require proactive methods in network management, where potentially severe issues could be predicted and avoided before they can impact data exchanges. Our ongoing research focusses on leveraging deep learning (DL) methodologies, particularly Transformer-based model, to analyse network paths, and explore inter-connectivity across networks with the goal of predicting key performance metrics.

A key challenge in network topology modelling is handling missing or uncertain path segments. To address this, we incorporate confidence-aware learning in our Transformer model. This approach enables a more effective representation of network paths, leading to improved model performance.

In this work, we present our experimental findings, discuss the challenges associated with incomplete network paths, and compare the performance of our model with baseline models. Our results demonstrate the potential of transformer-based models in the refinement of network analysis and pave the way to more robust topology-based anomaly detection methods.