LogHyper-IB: Spatio-Temporal Hypergraph Representation Learning via Information Bottleneck for Anomaly Detection in Distributed Systems

The operational integrity of High-Performance Computing (HPC) architectures like BlueGene/L is increasingly threatened by the “redundancy problem”, where critical failure signatures are submerged beneath ephemeral telemetry. This study introduces LogHyper-IB, a bifurcated architectural innovation that displaces traditional sequential paradigms by integrating Hypergraph Convolutional Networks (HGCN) with a Variational Information Bottleneck (VIB). By modeling log templates as nodes within a hyper-relational topology, the framework captures higher-order, non-pairwise dependencies often obscured in linear time-series. The VIB objective enforces a rigorous compression constraint, distilling “minimal sufficient statistics” by filtering out uninformative system heartbeats. Empirical validation on the BGL dataset yields a Matthews Correlation Coefficient (MCC) of 0.9888 and an F1-score of 0.9899, significantly outperforming legacy models. With a sub-50ms inference latency and robust root-cause localization capabilities, LogHyper-IB establishes a new performance ceiling for automated, real-time supercomputing diagnostics.