Impact of Distances in an Anomaly Detection Context for Time Series in Software Testing

Impact of Distances in an Anomaly Detection Context for Time Series in Software Testing

연구 분야: Verification

학회: International Conference on Computational Science and Its Applications

Software Testing is a critical stage in Software Development projects, which increases their quality at the cost of an increasing budget. This cost grows as the System Under Test (SUT) has poor testability. Data streaming systems or web servers usually present poor testability because they have constant input rates that often hide faults during testing. For such systems, it is costly to build test cases and verify whether the output is correct for each particular input. Tricorder proposes a methodology for anomaly detection without system requirements, taking into account the unsupervised learning provided by Damicore’s methodology. This paper evaluates the behavior of distance measures for time series data within the frameworks of unsupervised machine learning carried out by Damicore, which is applied to software anomaly detection. The study systematically assesses the impact of each distance measure on the resulting accuracy within Damicore by conducting experiments across multiple benchmarks. Our findings reveal that the Levenshtein distance significantly outperforms DTW, NCD, FFT, and Hamming. Conversely, Hamming distance demonstrates the poorest performance. Levenshtein distance, on the other hand, offers an excellent balance between execution time and accuracy, providing an optimal trade-off for efficient and effective fault detection. The findings provide valuable insights into the effectiveness of these measures in enhancing fault detection accuracy, which could be extended to other domains involving anomaly detection of time-series data.