Abstract:
Abstract: Log-oriented anomalous event prediction provides important support for security diagnosis， intelligent operation and maintenance of complex systems. Most of the existing mainstream anomalous event prediction methods based on deep learning technology capture sequence features from the local perspective of event sequence segments and the feature types are relatively single， resulting in low prediction accuracy. To solve this problem， an anomalous event prediction method based on graph neural network is proposed. The log sequence is represented as a graph structure with log events as nodes and the relationship between events as edges， so that it can simultaneously depict the log sequence from the perspective of semantics， statistics and event relationship to capture its spatio-temporal dynamic characteristics to improve the prediction performance. On this basis， the anomaly prediction task is transformed into a graph classification problem， and an anomaly prediction model based on graph isomorphic network is established by training graph neural network， which can more accurately capture the difference between the log sequence before the failure and the log sequence under normal conditions， and further improve the performance of anomaly prediction. The experimental results on three benchmark datasets show that the average F1 of the proposed method is 0.958， which is better than the comparison methods， and it can accurately predict anomalous events for early warning.

Key words: Key words: log parsing, anomaly prediction, graph neural network, event mining