Anomaly and Novelty Detection
Detection targets and supervision
- Outlier versus novelty detection - Outlier detection fits potentially contaminated observations; novelty detection learns a reference population and scores new inputs. The training contamination and availability of labels define different tasks. docs.
- Anomaly structure - Unusual points, unusual values conditional on context, and unusual sequences or groups require different scoring rules. Ruff et al. organize methods by assumptions about normality and available supervision. Ruff et al. 2021.
- Scores versus alarms - An anomaly ranking does not specify an operational threshold or a probability of failure. Detection interfaces separate scores from thresholded predictions; contamination settings encode a cutoff choice. docs.
Neighborhoods, isolation and distribution tails
- LOF - Compares local reachability density with neighboring densities, detecting points sparse relative to their neighborhood. Neighborhood size changes the scale of abnormality. Breunig et al. 2000.
- Isolation Forest - Random feature/split partitions isolate observations; short average isolation paths indicate anomalies. Isolation replaces explicit density or distance estimation. Liu et al. 2008.
- Extended Isolation Forest - Uses randomly oriented hyperplanes instead of axis-aligned splits to reduce score artifacts. The geometric change addresses partition bias rather than learning semantic anomaly labels. Hariri et al. 2021.
- ECOD - Estimates empirical marginal cumulative distributions, converts observations to tail scores and aggregates across features. Its featurewise construction supplies a simple tabular baseline without learning arbitrary joint dependencies. Li et al. 2023.
- COPOD - Constructs an empirical copula and combines tail information to score extreme observations. Distribution-tail scoring differs from LOF’s comparison against a local neighborhood. Li et al. 2020.
Support and robust distribution models
- One-class SVM - Learns a function separating an estimated support region from its complement using kernels. Kernel and regularization choices define the boundary; it is not a labeled anomaly classifier. Schölkopf et al. 2001.
- Robust covariance - Minimum covariance determinant / EllipticEnvelope fit a robust location and covariance for distance-based detection. Useful when an approximately elliptical normal population is plausible; multimodal structure needs other models. Rousseeuw and Van Driessen 1999; docs.
Learned one-class and reconstruction methods
- Deep SVDD - Learns a representation concentrated near a hypersphere center. Architecture and training restrictions address trivial collapse; a compact representation alone is not evidence of useful anomaly separation. Ruff et al. 2018.
- Deep SAD - Extends deep one-class learning to unlabeled data plus labeled normal or anomalous examples. Its label-informed setting differs from training only on presumed normal data. Ruff et al. 2020.
- DROCC - Trains against adversarially generated nearby negatives under a local low-dimensional-manifold assumption. The generated contrast examples define the learned one-class boundary. Goyal et al. 2020.
- Reconstruction and MemAE - Ordinary autoencoders can reconstruct anomalies well. Gong et al. restrict reconstruction through a learned memory of normal patterns; low reconstruction error is not inherently a normality guarantee. Gong et al. 2019.
Classifier-based and out-of-distribution detection
- Maximum softmax probability - Uses classifier confidence as a baseline for misclassification and out-of-distribution detection. This needs a trained classifier and does not turn confidence into a novelty guarantee. Hendrycks and Gimpel 2017.
- Outlier Exposure - Trains detectors using auxiliary outlier data to improve detection of unseen anomalies. Auxiliary-data choice and overlap with evaluated anomalies are part of the supervision protocol. Hendrycks et al. 2019.
- Likelihood failure - Tested flows, VAEs and PixelCNNs trained on CIFAR-10 can assign higher likelihood to SVHN images. High generative density is not a general test of semantic in-distribution membership. Nalisnick et al. 2019.
- Likelihood-ratio OOD detection - Compares a generative model with a background model to reduce nuisance-statistic effects. Ren et al. improve detection on genomic and image tasks; the result depends on what the background model removes, not simply on using a density model. Ren et al. 2019.
Sequences and streaming thresholds
- Bayesian online changepoint detection - Maintains a posterior over time since the last change using a predictive model and hazard function. Detects changes in a sequence’s generating regime; an isolated unusual observation need not imply a persistent regime change. Adams and MacKay 2007.
- Matrix-profile discords - Rank fixed-length subsequences by distance to their nearest nontrivial match. Window length and normalization define the anomaly; repeated faults may cease to be rare patterns. Yeh et al. 2016.
- OmniAnomaly - Uses stochastic recurrent representations and normalizing flows, scoring multivariate sequences by reconstruction probabilities. Per-channel scores localize discrepancies rather than identify causal fault sources. Su et al. 2019.
- Anomaly Transformer - Contrasts learned temporal attention with a locality prior using association discrepancy. Its anomaly criterion assumes abnormal points have different temporal associations; this is a modeling hypothesis. Xu et al. 2021.
- TranAD - Combines transformer reconstruction, self-conditioning and adversarial training for multivariate time series. Detection scores and diagnostic rankings need evaluation under the chosen temporal and threshold protocol. Tuli et al. 2022.
- Extreme-value thresholds - Siffer et al. adapt streaming thresholds from tail behavior and a specified risk parameter. Threshold estimation is separate from learning representations; tail-model fit matters. Siffer et al. 2017.
Visual anomaly detection and localization
- PaDiM - Fits multivariate Gaussian distributions to pretrained patch embeddings to detect and localize deviations from normal images. The reference distribution is spatially indexed rather than one global image density. Defard et al. 2020.
- PatchCore - Compares pretrained patch features with a representative memory bank of normal patches. Nearest-reference scoring supports image detection and local defect maps without training a classifier on defect classes. Roth et al. 2022.
- WinCLIP - Combines text prompts with window, patch and image features for zero-shot anomaly detection and segmentation. WinCLIP+ additionally uses normal reference images; these are different data-access regimes. Jeong et al. 2023.
Comparative evidence and evaluation
- ADBench - Compares 30 detectors on 57 datasets across supervision, anomaly types and corruption. No tested unsupervised method dominates statistically across datasets; strong methods depend on anomaly structure. Han et al. 2022.
- Label value in ADBench - Revealing 1% of anomaly labels lets most tested semi-supervised methods beat the best unsupervised baseline in that study. This is a fraction of anomalies, not 1% of all observations. Han et al. 2022.
- Point adjustment - Crediting an entire true anomaly segment after one detected point can inflate F1, even for random scores. Report unadjusted scores and the exact event-credit rule. Kim et al. 2021.
- Dataset artifacts - Wu and Keogh identify labeling and construction flaws in widely used time-series benchmarks and introduce the UCR anomaly archive. Their criticism concerns examined datasets; it does not invalidate all temporal detection. Wu and Keogh 2021.
- MVTec AD 2 - Adds eight industrial scenarios, including small defects and lighting changes. Heckler-Kram et al. report tested methods below 60% average AU-PRO; earlier near-saturated benchmarks conceal these difficulties. Heckler-Kram et al. 2026.
- Rare-event metrics - Report class prevalence, precision-recall performance and false alarms at a threshold chosen without test labels. High ROC-AUC alone does not specify an acceptable operational alert rate. Davis and Goadrich 2006.
- Temporal evaluation protocol (editorial guidance) - Separate point detection, event detection and delay; keep device/time splits intact and state whether future samples are available. Evaluation critique motivates reporting the rule that turns scores into credited events. Kim et al. 2021.
Software
- scikit-learn - Outlier detection includes Isolation Forest, LOF, one-class SVM and robust covariance. LOF’s novelty mode scores new observations; its training-score behavior differs from ordinary outlier mode. docs.
- PyOD - Provides a common Python interface for classical, ensemble and neural outlier detectors. Shared interfaces aid comparisons but do not standardize preprocessing, tuning or label access automatically. Zhao et al. 2019.
- Anomalib - Visual anomaly-detection toolkit with model implementations, data pipelines and evaluation support. Record backbone, preprocessing and reference-image access when comparing methods. code.
Surveys
- Unifying review - Ruff et al. connect density, support estimation and reconstruction approaches, including shallow/deep relationships. Their organizing assumptions help explain why the same score behaves differently across tasks. Ruff et al. 2021.
Star
Issue