DiffInt

Differentiable Interval Bottlenecks for Interpretable Anomaly Detection in Numerical Data.

DiffInt is a reconstruction autoencoder whose latent bottleneck is a set of soft, axis-aligned interval memberships learned end-to-end from raw numerical data (no discretization). Each latent unit is a learnable hyper-rectangle (feature j ∈ [a, b]); an instance is encoded by how strongly it falls inside each interval, and the per-sample reconstruction error is the anomaly score. The interval structure makes the bottleneck directly inspectable, so a flagged point comes with feature-range explanations rather than a single opaque score.

Key properties:

• Interpretable by design: the bottleneck is a set of readable intervals.
• A deterministic guarantee: out-of-support points are clipped to a fixed reconstruction (the clipping-margin property), which is why MAE is a valid score.
• Required [-1,1] normalization: calibrates initialization, the sigmoid temperature, and the guarantee; one configuration transfers across datasets.
• Label-free importance (LFI): a closed-form score per (unit, feature) computed from quantities the trained model already keeps, used to surface candidate constraints without anomaly labels.

Supplementary material

supplementary.pdf accompanies the paper. It collects the material that does not fit in the main text: dataset characteristics, the complete per-dataset ROC–AUC and AUPRC tables (48 datasets × 23 methods, mean ± std with ranks), per-scale critical-difference diagrams on both metrics, the per-dataset robustness ablations (normalization, contamination, τ sensitivity, active-coordinate threshold, inference clamping, LFI seed stability, per-layer spectral norms), the complexity/scalability profile, and the LFI faithfulness study.

Install

git clone https://github.com/DiffInt/diffint
cd diffint
pip install -e .                      # or: pip install -r requirements.txt

Requires Python ≥ 3.8 and PyTorch (CPU is fine). Plotting interval-level explanations additionally needs seaborn/matplotlib (pip install -e ".[explain]").

Quickstart

import numpy as np
from sklearn.metrics import roc_auc_score
from diffint import DiffInt

# X_train: inliers only (semi-supervised). X_test/y_test: mixed.
det = DiffInt(K=200, epochs=1000).fit(X_train)
scores = det.decision_function(X_test)        # higher = more anomalous
print("AUROC", roc_auc_score(y_test, scores))

Inputs are min–max scaled to [-1, 1] internally (and clipped to [-1, 1] at inference); do not pre-scale. See examples/quickstart.py.

Reading the intervals

for unit, feat, lo, hi in det.interval_rules(feature_names=cols)[:10]:
    print(f"unit {unit}: {feat} in [{lo:.2f}, {hi:.2f}]")

Reproduce the ADBench evaluation

Place ADBench CSVs (feature columns + final 0/1 label column) in csv/:

python scripts/run_adbench.py --data_dir csv/ --K 200 \
    --epochs 1000 --lr 5e-5 --nb_runs 3

Protocol: semi-supervised (train on inliers), [-1,1] normalization for all methods, 40% test split, mean AUROC/AUPRC over seeds.

Package layout

diffint/
  estimator.py               # DiffInt: fit / decision_function / predict / interval_rules
  autoencoder.py             # interval-bottleneck autoencoder (+ optional spectral norm)
  intervalPatternNetwork.py  # soft memberships + pattern aggregator
  intervalparam.py           # interval centers + softplus half-widths
  utils.py                   # batched scoring helpers
scripts/run_adbench.py       # ADBench reproduction
examples/quickstart.py       # minimal example

License

MIT — see LICENSE.