// the find
KDD-OpenSource/DeepADoTS
Repository of the paper "A Systematic Evaluation of Deep Anomaly Detection Methods for Time Series".
A research benchmark comparing seven deep learning anomaly detection methods on time series data, originally accompanying a KDD paper. It implements LSTM-AD, LSTM-ED, Autoencoder, Donut, REBM, DAGMM, and LSTM-DAGMM under a unified scikit-learn-style API so you can run apples-to-apples comparisons. Aimed at researchers who want a reproducible baseline, not practitioners building production systems.
Unified fit/predict interface across all seven algorithms makes switching between methods a one-line change. Includes both synthetic and real-world dataset generators, so you're not stuck wiring in your own data pipeline before seeing results. The evaluation harness (ROC AUC, plotter, per-method reports) is already there — you don't have to build the comparison scaffolding yourself. Docker support means the dependency hell of mixing TF1-era and PyTorch code is someone else's problem.
The codebase is frozen in TF1/TF2-early-days territory — Donut and REBM depend on TensorFlow patterns that have changed substantially, and the requirements.txt almost certainly won't install cleanly against modern Python without pinning ancient versions. The paper is from 2018-2019, so none of the transformer-based or diffusion-based anomaly detection methods from the last five years are represented. Test coverage is minimal (one initialization test file), so you're taking the correctness of each implementation largely on faith. No pretrained weights or saved checkpoints — every experiment requires a full training run, which is painful when DAGMM can take hours on real datasets.