On the Influence of Data Resampling for Deep Learning-Based Log Anomaly Detection: Insights and Recommendations

Numerous DL-based approaches have garnered considerable attention in the field of software Log Anomaly Detection. However, a practical challenge persists: the class imbalance in the public data commonly used to train the DL models. This imbalance is characterized by a substantial disparity in the number of abnormal log sequences compared to normal ones, for example, anomalies represent less than 1% of one of the most popular datasets. Previous research has indicated that existing DLLAD approaches may exhibit unsatisfactory performance, particularly when confronted with datasets featuring severe class imbalances. Mitigating class imbalance through data resampling has proven effective for other software engineering tasks, however, it has been unexplored for LAD thus far. This study aims to fill this gap by providing an in-depth analysis of the impact of diverse data resampling methods on existing DLLAD approaches from two distinct perspectives. Firstly, we assess the performance of these DLLAD approaches across three datasets and explore the impact of resampling ratios of normal to abnormal data on ten data resampling methods. Secondly, we evaluate the effectiveness of the data resampling methods when utilizing optimal resampling ratios of normal to abnormal data. Our findings indicate that oversampling methods generally outperform undersampling and hybrid methods. Data resampling on raw data yields superior results compared to data resampling in the feature space. In most cases, certain undersampling and hybrid methods show limited effectiveness. Additionally, by exploring the resampling ratio of normal to abnormal data, we suggest generating more data for minority classes through oversampling while removing less data from majority classes through undersampling. In conclusion, our study provides valuable insights into the intricate relationship between data resampling methods and DLLAD.