xLSTMAD: A Powerful xLSTM-based Method for Anomaly Detection [ICDM 2025]

This repository contains the improved official implementation of xLSTMAD, the first anomaly detection framework built upon a full encoder–decoder xLSTM architecture for multivariate time series (see paper). xLSTMAD was published at the ICDM 2025 (IEEE International Conference on Data Mining).

List of changes in comparison to the original version:

• We improved the implementation to avoid bottlenecks, making xLSTMAD much faster.
• We leverage PyTorch Lightning to make it as easy to use as possible.
• For simplicity, this version focuses only on the reconstruction variant and the MSE error.

The original implementation, along with the experimental setup for TSB-AD-M is available here.

The paper is available here: https://arxiv.org/pdf/2506.22837. When using, please cite:

   @INPROCEEDINGS{xlstmad,
    author={Faber, Kamil and Pietron, Marcin and Zurek, Dominik and Corizzo, Roberto},
    booktitle={2025 IEEE International Conference on Data Mining (ICDM)},
    title={xLSTMAD: A Powerful xLSTM-based Method for Anomaly Detection},
    year={2025},
    volume={},
    number={},
    pages={247-256},
    doi={10.1109/ICDM65498.2025.00032}}

Installation

1. Create python environment or use an existing one (we tested xLSTMAD with Python 3.11)
2. Download pytorch and torchvision from here and install them according to your system configuration.
3. Install required dependencies by running:

pip install -r requirements.txt

Example usage

The main code of xLSTMAD method is in xlstmad.py file. Our example script trains xLSTMAD on the example dataset and evaluates it on the test set. Its main goal is to showcase how to use the xLSTMAD class (pytorch lightning module) and how to prepare the data for it.

You can run it by executing the following command: python example.py

The script accepts just one parameter --slstmbackend that determines which backend to use for the SLSTM cell. The available options are cuda and vanilla. The cuda backend is much faster but requires NVIDIA Compute Capability ≥ 8.0. The vanilla backend is slower but can run on any system without additional dependencies.