Unsupervised Model Selection for Time Series Anomaly Detection

Authors: Mononito Goswami, Cristian Ignacio Challu, Laurent Callot, Lenon Minorics, Andrey Kan

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Large-scale experiments on multiple real-world datasets demonstrate that our proposed unsupervised approach is as effective as selecting the most accurate model based on partially labeled data.
Researcher Affiliation Collaboration Mononito Goswami & Cristian Challu Carnegie Mellon University {mgoswami,cchallu}@cs.cmu.edu Laurent Callot, Lenon Minorics & Andrey Kan Amazon Research {lcallot,avkan}@amazon.com, minorics@amazon.de
Pseudocode No The paper describes its methods in prose and explains algorithms conceptually but does not include any formal pseudocode blocks or clearly labeled algorithm figures.
Open Source Code Yes We provide an open-source implementations of our model selection method and anomaly detection models at https://github.com/mononitogoswami/tsad-model-selection.
Open Datasets Yes All datasets used in this work are available in the public domain. Directions to download the Server Machine Dataset are available at https://github.com/Net Man AIOps/Omni Anomaly, whereas the UCR Anomaly Archive can be downloaded from https://www.cs.ucr.edu/ eamonn/time_series_data_2018/.
Dataset Splits Yes Specifically, we randomly partition each dataset into selection (20%) and evaluation (80%) sets. Each time-series in the dataset is assigned to one of these sets. The baseline method selects a model based on anomaly labels in the selection set.
Hardware Specification Yes All our experiments were carried out on an AWS g3.4xlarge EC2 instance with with 16 Intel(R) Xeon(R) CPUs, 122 Gi B RAM and a 8 Gi B GPU.
Software Dependencies Yes All models and model selection algorithms were trained and built using scikit-learn 1.1.1 (Pedregosa et al., 2011), Py Torch 1.11.0 Paszke et al. (2019) along with Python 3.9.13.
Experiment Setup Yes We created a pool of 3 k-NN (Ramaswamy et al., 2000), 4 moving average, 4 DGHL (Challu et al., 2022), 4 LSTMVAE (Park et al., 2018) and 4 RNN (Chang et al., 2017) models by varying hyperparameters for each model, for a total 19 combinations. Finally, to efficiently train our models, we sub-sampled all time-series with T > 2560 by a factor of 10.