Running summarizing tools on a new article
Check the paper below, then start its synthesis.
Deep Anomaly Detection in Unstructured System Log Files with Ensemble-Based Learning
Mikhail E. Zuev (Saint Petersburg State Electrotechnical University), Artyom D. Semidolin (Saint Petersburg State Electrotechnical University)
Abstract: Anomaly detection in unstructured system logs remains challenging due to high variance of deep learning models across training subsamples, leading to unstable performance. While BERT-based approaches like LogBERT achieve state-of-the-art results, their sensitivity to data shifts limits reliability. To address this, we propose an ensemble of LogBERT models using bootstrap aggregating (bagging) to reduce prediction variance and improve robustness. The implemented method trains multiple LogBERT instances on bootstrapped samples of normal log sequences and combines their outputs via majority voting. Evaluated on standard LogHub datasets — HDFS, BGL, Thunderbird (first 20M log lines), and Spirit (almost 8M log lines) — the ensemble consistently outperforms single models. We achieve F1-scores of 86.43% (HDFS), 93.34% (BGL), 100% (Thunderbird-mini), and 100% (Spirit), surpassing baseline LogBERT by 6.87%, 3.23%, 3.37%, and 1.15%, respectively. The average improvement is 3.95%, demonstrating that bagging significantly enhances accuracy and stability in log anomaly detection, especially for imbalanced and heterogeneous log data.