TAD 벤치마크 데이터셋

• 최신 6개의 논문에서 사용된 Time Series 데이터셋 정리 및 모델 간의 성능 비교를 위해 작성
• 각 모델별 사용한 데이터의 특성, 모델의 특성이 다르기 때문에 해당 표는 단순히 참고용
• 논문 목록
  • Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy
    • Reconstruction-based method + Association Discrepancy loss를 추가
  • Time Series Anomaly Detection with Multi-resolution Ensemble Decoding
    • Reconstruction-based method + Multi-resolution 기법을 추가
  • Detecting Anomalous Event Sequences with Temporal Point Processes
    • Temporal point precess에서의 OOD detection을 AD로 해석
  • Neural Contextual Anomaly Detection for Time Series
    • Window 기법
  • Multi-Scale One-Class Recurrent Neural Networks for Discrete Event Sequence Anomaly Detection
    • OC-NN, event sequence + local, global latent space 를 따로 지정해 OC-NN을 수행
    • 로그 메시지를 미리 설정하여 이상현상을 찾아내는 모델
  • Multivariate Time Series Anomaly Detection and Interpretation using Hierarchical Inter-Metric and Temporal Embedding
    • 다변량 시계열 데이터를 2가지 관점에서 해석하여 각각을 임베딩하여 학습하는 모델
      • 하나의 변수 내에서 시간적인 변화량 관점
      • 여러 변수간의 상관관계 관점

데이터셋

(1) SMD (Server Machine Dataset, Su et al. (2019))

• 5-week-long dataset that is collected from a large Internet company with 38 dimensions.
• https://github.com/NetManAIOps/OmniAnomaly/tree/master/ServerMachineDataset
• https://gudwns1215.medium.com/how-to-time-series-anomaly-detection-using-deep-learning-edfc2d1dbbc8

(2) PSM (Pooled Server Metrics, Abdulaal et al. (2021))

• Multiple application server nodes at eBay with 26 dimensions.
• https://github.com/eBay/RANSynCoders/tree/main/data

(3) Both MSL (Mars Science Laboratoryrover) and SMAP (Soil Moisture Active Passive satellite)

• Public datasets from NASA (Hundman et al., 2018) with 55 and 25 dimensions respectively,
• Contain the telemetry anomaly data derived from the Incident Surprise Anomaly (ISA) reports of spacecraft monitoring systems.
• 데이터 확보가 어려워 일단 제외

(4) SWaT(Secure Water Treatment, Mathur & Tippenhauer (2016))

• Obtained from 51 sensors of the critical infrastructure system under continuous operations.
• https://itrust.sutd.edu.sg/itrust-labs_datasets/

(5) NeurIPS-TS (NeurIPS 2021 Time Series Benchmark)

• Proposed by Lai et al. (2021) and includes five time series anomaly scenarios categorized by behavior-driven taxonomy as point-global, pattern-contextual, pattern-shapelet, pattern-seasonal and pattern-trend.

(6) UCR Dataset

• Multi-dataset Time Series Anomaly Detection Competition of KDD2021.

(7) ECG

• This is a collection of 6 data sets on the detection of anomalous beats from electrocardiograms readings.
• https://www.cs.ucr.edu/~eamonn/discords/

(8) 2D-gesture

• This contains time series of X-Y coordinates of an actor’s right hand. The data is extracted from an video in which the actor grabs a gun from his hip-mounted holster, moves it to the target, and returns it to the holster. The anomalous region is in the area where the actor failsto return his gun to the holster.
• https://www.cs.ucr.edu/~eamonn/discords/

(9) Power-demand

• This contains one year of power consumption records measured by a Dutch research facility in 1997.
• https://www.cs.ucr.edu/~eamonn/discords/

(10) Yahoo’s S5 Webscope

• This contains records from real production traffic of the Yahoo website
• https://webscope.sandbox.yahoo.com/catalog.php?datatype=s

(11) RUBiS

• weblog dataset and was generated by an auction site prototype modeled after eBay.com
• User web behavior including user_id, date, request_inf

(12) Hadoop Distributed File Systme(HDFS)

• Generated by a Hadoop-based map-reduce cloud environment usingbenchmark workloads
• 11,175,629 log messages, of which 2.9% are labeled as anomalies by Hadoop experts

(13) BlueGene/L (BGL)

• 4,747,936 log messages generated by a BlueGene/L supercomputer system with 131,072 processors and 32,768 GB memory at Lawrence Livermore National Labs.

(14) WADI

• Water Distribution

(15) ASD

• Application Server Dataset
• https://github.com/zhhlee/InterFusion/tree/main/data