Detecting Multivariate Time Series Anomalies with Zero Known Label
Authors: Qihang Zhou, Jiming Chen, Haoyu Liu, Shibo He, Wenchao Meng
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on five public datasets with seven baselines are conducted, MTGFlow outperforms the SOTA methods by up to 5.0 AUROC%. |
| Researcher Affiliation | Collaboration | Qihang Zhou1, Jiming Chen1, Haoyu Liu1,2*, Shibo He1,3, Wenchao Meng1 1Zhejiang University, Hangzhou, China, 2Net Ease Fuxi AI Lab, Hangzhou, China, 3Key Laboratory of Collaborative Sensing and Autonomous Unmanned Systems of Zhejiang Province, Hangzhou, China |
| Pseudocode | No | The paper does not contain pseudocode or an algorithm block. |
| Open Source Code | Yes | Code is available at github.com/zqhang/MTGFLOW. |
| Open Datasets | Yes | The commonly used public datasets for MTS anomaly detection in OCC are MSL (Mars Science Laboratory rover) (Hundman et al. 2018), SMD (Server Machine Dataset) (Su et al. 2019), PSM (Pooled Server Metrics) (Abdulaal, Liu, and Lancewicki 2021), SWa T (Secure Water Treatment) (Goh et al. 2016) and WADI (Water Distribution) (Ahmed, Palleti, and Mathur 2017). |
| Dataset Splits | Yes | For other datasets, the training split contains 60% data, and the test split contains 40% data. For SWa T, the training split contains 60% data, the validation split contains 20% data, and the test split contains 20% data. |
| Hardware Specification | Yes | The epoch is 40 for all experiments, which are performed in Py Torch1.7.1 with a single NVIDIA RTX 3090 24GB GPU1. |
| Software Dependencies | Yes | The epoch is 40 for all experiments, which are performed in Py Torch1.7.1 with a single NVIDIA RTX 3090 24GB GPU1. |
| Experiment Setup | Yes | For all datasets, we set the window size as 60 and the stride size as 10. Adam optimizer with a learning rate 0.002 is utilized to update all parameters. One layer of LSTM is sufficient to extract time representations in our experiment. One self-attention layer with 0.2 dropout ratio is adopted to learn the graph structure. We use MAF as the normalizing flow model. For SWa T, one flow block and 512 batch size are employed. For other datasets, we arrange two flow blocks for it and set the batch size as 256. λ is set as 0.8 for thresholds of all entities. The epoch is 40 for all experiments |