Timeseries Anomaly Detection using Temporal Hierarchical One-Class Network

Authors: Lifeng Shen, Zhuocong Li, James Kwok

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive empirical studies on various real-world timeseries demonstrate that the proposed THOC network outperforms recent strong deep learning baselines on timeseries anomaly detection. Experiments performed on a number of real-world timeseries data sets show that the proposed model outperforms the recent state-of-the-arts.
Researcher Affiliation Collaboration Lifeng Shen1, Zhuocong Li2, James T. Kwok1 1 Department of Computer Science and Engineering, Hong Kong University of Science and Technology, Hong Kong {lshenae,jamesk}@cse.ust.hk 2 Cloud and Smart Industries Group, Tencent, China zhuocongli@tencent.com
Pseudocode Yes Algorithm 1 Temporal hierarchical one-class learning (THOC). Input: timeseries Xs = (x1,s, x2,s, . . . , x Ts,s); number of centers {Kl}; skip lengths {s(l)}. 1: repeat 2: feed xt,s into the L-layer dilated-RNN and obtain {f l t} from each layer; 3: for layer l = 1, . . . , L do 4: obtain the lth clustering layer s input { f l 1 t,i }i=1,...,Kl 1 by (8) where K0 =1; 5: compute probabilities {P l t,i j}i=1,...,Kl 1,j=1,...,Kl from (5); 6: compute {Rl t,j}j=1,...,Kl for each cluster center at layer l from (10); 7: update and obtain output features {ˆf l t,j}j=1,...,Kl from (7); 8: end for 9: minimize MVDD objective in (13) by the Adam optimizer; 10: until convergence.
Open Source Code No The paper states 'Source codes of the baselines are downloaded from the web.' but does not provide a statement or link for the source code of their own proposed method.
Open Datasets Yes The following timeseries data sets are used: (i) 2D-gesture [9], which records the X-Y coordinate sequences of hand gestures in a video; (ii) Power demand[9], which contains a year of power demand at a Dutch research facility; (iii) KDD-Cup99 data from the DARPA 98 Intrusion Detection Evaluation Program [14]. It contains around seven million network traffic connection records over a 7-week period. (iv) Secure Water Treatment (SWa T) data [18], which is collected from a water treatment testbed over 11 days. (v) Mars Science Laboratory rover (MSL); and (vi) Soil Moisture Active Passive satellite (SMAP) data: Both MSL and SMAP are public data sets from NASA [8].
Dataset Splits Yes For 2D-gesture, power-demand, KDD-Cup99, and SWa T, the raw data set has only a training set and a test set. To allow model selection and hyperparameter tuning, we use part of the provided test set for validation. For MSL and SMAP, we follow the setting in [28], and hold out 30% of the training data as validation set. A summary of the resultant data set statistics is shown in Table 1.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts, or cloud instance types) used for running its experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., library names with versions).
Experiment Setup No The paper states 'Detailed experimental settings can be found in Appendix B.', implying that specific setup details like hyperparameters are not present in the main text provided.