Answers are grounded in this paper and cite their sources.
Log-based anomaly detection is crucial for ensuring the reliability of large-scale distributed systems. DeepLog implemented a sequence-oriented methodology utilizing LSTMs, providing a data-driven substitute for rule-based techniques. Nonetheless, it is constrained in its capacity to capture contextual relationships, temporal dynamics, and semantic shifts within log sessions. This study introduces a multi-faceted extension of DeepLog that incorporates diverse viewpoints to enhance detection precision and resilience. A BiLSTM with attention models symbolic log key sequences, whereas a parameter-based regression model captures inter-event patterns via temporal prediction. The two perspectives are integrated via an XGBoost classifier trained on soft anomaly scores. The complete pipeline is refined using Bayesian hyperparameter optimization using Optuna, guaranteeing uniform generalization. Experiments conducted on the HDFS dataset from LogHub demonstrate significant enhancements in F1-Score, AUC, and False Positive Rate relative to the original DeepLog. These findings underscore the significance of integrating symbolic, temporal, and statistical cues to improve log anomaly identification. Evaluated on the HDFS benchmark, FuseLog achieves an F1-Score of 0.980 and a Recall of 0.999. While results are promising, the current validation is limited to identifier-driven log formats, necessitating future adaptation for stream-based environments.
No takes yet. Share an insight, caveat, or question.
Alzamil et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: