Temporal Evidence Aggregation with Causal-Contrastive Transformer Learning for Intelligent Cloud Failure Diagnosis

Abstract

Addressing the challenges of job failures in distributed training clusters, characterized by cross-layer coupling, heterogeneous evidence, and temporal propagation, this paper investigates the problem of job failure cause identification and proposes a unified modeling framework based on Transformer. The method aligns indicator sequences and log events within the same time window, constructing a shared semantic representation through linear mapping and position injection, enabling the fusion of continuous resource fluctuations and discrete trigger signals in the same representation space. Subsequently, global self-attention is employed to model long-range dependencies, capturing the correlation between failure precursors and termination manifestations. Causal constraints are used to suppress unreasonable information leakage and strengthen the consistency of the inference chain. To improve the structural stability of class boundaries, contrastive constraints are introduced to enhance intra-class compactness and inter-class separation, thus maintaining discriminative reliability under noisy and missing conditions. Cause identification samples are constructed based on publicly available cluster tracking data and evaluated. Results show that the proposed method achieves superior performance in accuracy and ranking quality, providing more reliable cause identification outputs for cluster-level automated diagnosis and operational decision-making.