PERTANIKA JOURNAL OF SCIENCE AND TECHNOLOGY

In the industrial domain, logs are widely applied in the management and maintenance of software systems to ensure reliability and availability. Furthermore, in the research field, various deep learning methods such as CNNs, LSTMs, and Transformers have been reported to achieve high accuracy in anomaly detection studies. However, there are challenges to their adoption in development fields. One reason is the limited datasets used in research, which lack a comprehensive evaluation for general applicability. To address this, we have prepared metrics to assess the complexity of log datasets necessary for creating a log generator for research purposes. We conducted a comparative study on the complexity of datasets in both research and industrial domains. Our evaluation of log sequence complexity, using frequency of occurrence and the Gini coefficient, showed that industrial logs are more complex across all metrics. This highlights the increased need for datasets close to the industrial domain for research purposes. Our study's findings suggest that a clear metric for dataset complexity can be achieved by converting logs into templates and then into sequences of size 10, evaluated using the Gini coefficient or kurtosis. Future work will involve developing a generator that produces logs close to those found in development environments, using these metrics as target values.