WSAD-DT — Weakly Supervised Anomaly Detection via Dual-Tailed Kernel

Official implementation of the ICML 2025 paper "Weakly Supervised Anomaly Detection via Dual-Tailed Kernel" (Walid Durani, Tobias Nitzl, Claudia Plant, Christian Böhm).

WSAD-DT learns latent representations that separate anomalies from normal samples using only a handful of labeled anomalies. It introduces two centroids — one normal, one anomalous — and a dual-tailed kernel scheme: a light-tailed kernel compactly models in-class points while a heavy-tailed kernel maintains a wide margin against out-of-class instances; a kernel-based regularizer preserves intra-class diversity. An ensemble partitions the normal data across members (all members share the labeled anomalies), improving robustness.

Install

pip install wsad-dt

or from source:

git clone https://github.com/Walid10010/Weakly-Supervised-Anomaly-Detection-via-Dual-Tailed-Kernel.git
cd Weakly-Supervised-Anomaly-Detection-via-Dual-Tailed-Kernel
pip install -e .

Runtime dependencies: numpy, scikit-learn, torch (CPU is sufficient).

Quickstart

from wsad_dt import WSADDT

# y_weak: 1 for the few labeled anomalies, 0 for everything else
det = WSADDT(n_ensemble=5, seed=100).fit(X_train, y_weak)

scores = det.decision_function(X_test)   # higher = more anomalous
labels = det.predict(X_test)             # 1 = anomaly

Inputs should be scaled (the paper protocol uses MinMax to [0, 1]). Key parameters: n_ensemble=5 (paper's num_splits), seed=100 (member j trains with seed seed * (j+1)), batch_size=64. The paper's train(num_splits, X, y_semi, s) / test(...) entry points remain importable from wsad_dt for script-level use.

Reproducing the paper

Experiments use the ADBench datasets and compare against DeepSAD, DevNet, RoSAS, PReNet, GANomaly, and XGBOD:

pip install -r requirements-experiments.txt
python experiments/run_exp.py

Protocol: MinMax scaling, stratified 70/30 splits (random_state ∈ {1, 5, 10}), 5% of anomalies labeled (minimum 5), ensemble of 5, AUC-ROC and AUC-PR averaged over seeds and splits.

Implementation notes

The packaged model is verified equivalent to the original reference script (tests/test_wsaddt.py::test_equivalent_to_legacy_reference), with training running deterministically on CPU. Scoring convention follows PyOD: higher decision_function values indicate anomalies (score_samples gives the scikit-learn sign convention). See IMPLEMENTATION_NOTES.md for documented reference semantics.

Citation

@inproceedings{durani2025wsaddt,
  title     = {Weakly Supervised Anomaly Detection via Dual-Tailed Kernel},
  author    = {Durani, Walid and Nitzl, Tobias and Plant, Claudia and B{\"o}hm, Christian},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning},
  series    = {Proceedings of Machine Learning Research},
  volume    = {267},
  pages     = {14833--14866},
  publisher = {PMLR},
  year      = {2025}
}

License

Released under the MIT License.