The CompositeModel architecture (named as AE-FAR or VAE-FAR) includes a combination of an Autoencoder (AE) or Variational Autoencoder (VAE) and MSE Feedback with Attention and RNN(FAR). This architecture leverages error feedback to enhance model performance through a two-phase training process.
The training is structured in two main phases:
-
Pretraining Phase: The AE or VAE is trained independently to learn a compressed representation of the input data. This pretraining establishes a robust initialization for the CompositeModel.
-
CompositeModel Training Phase (AE-FAR/VAE-FAR): The pretrained AE or VAE model is integrated into the CompositeModel with the FAR module. This phase refines the model to incorporate error feedback, enhancing its predictive capability.
Performance comparison of different anomaly detection methods in the literature.
- models folder contain pretrained models for proposed AE-FAR
- train-model.ipynb for training code
- test-model.ipynb to evaluate the model with pretrained weights
If you find this repo useful, please cite our paper.
@inproceedings{delibasoglu2024time,
title={Time Series Anomaly Detection Leveraging MSE Feedback with AutoEncoder and RNN},
author={Delibasoglu, Ibrahim and Heintz, Fredrik},
booktitle={31st International Symposium on Temporal Representation and Reasoning (TIME 2024)},
pages={17--1},
year={2024},
organization={Schloss Dagstuhl--Leibniz-Zentrum f{\"u}r Informatik}
}