DSGN: Log-based anomaly diagnosis with dynamic semantic gate networks

Abstract

Existing log anomaly diagnosis methods still face challenges in the lack of statistical features of log messages and insufficient exploitation of textual semantic features. In order to tackle this issue, we propose a novel approach called Dynamic Semantic Gating Network (DSGN). The core idea of DSGN is to enrich the semantic representation of log texts by selectively utilizing statistical information, thus achieving an organic integration of statistical and semantic features. Specifically, DSGN incorporates a variational encoding module to encode statistical features, and a log content-aware graph convolutional network module to capture semantic features from the log context. Furthermore, DSGN introduces a dynamic semantic threshold mechanism that dynamically adjusts the information flow based on the confidence level of semantic features and feeds it into the classifier. This design not only helps train a more robust classifier, but also leverages the advantages of both statistical and semantic features while avoiding overfitting caused by using statistical features. Experimental results show that the DSGN model achieves significant performance improvements on seven public datasets, with a macro-average F1 score exceeding 83% and a micro-average F1 score exceeding 81%, outperforming existing baseline techniques and demonstrating its substantial advantages.