Unsupervised Representation Learning for Fault Diagnosis in IT Infrastructure

The escalating complexity and scale of IT systems have rendered single-source fault diagnosis ineffective. One key factor is the limited capacity of individual data sources to comprehensively encompass system anomalies. Moreover, the intricate nature of IT systems, in conjunction with a diverse spectrum of potential faults, necessitates a more comprehensive monitoring strategy for optimal diagnostic outcomes. Consequently, multi-source data fault diagnosis has emerged as the prevailing approach in the contemporary IT landscape. Nevertheless, the heterogeneity in monitoring targets, techniques, and recording methodologies across distinct data sources introduces significant disparities, presenting significant challenges for the effective extraction of features and the learning of representations from such diverse data. Therefore, extracting features effectively and learning representations comprehensively from multiple data sources play a significant role in proficient fault diagnosis over complex IT systems. Addressing this concern, we introduce HIT, an innovative unsupervised representation learning paradigm founded on hypergraphs, tailored explicitly for multi-source IT infrastructure monitoring data. The acquired representations distinctly enhance fault diagnosis efficacy in IT systems, yielding promising outcomes.