After normalization and image feature calculation, we performed unsupervised outlier detection on Datasets A, B, C, A*, and B*, using 26 different categories of algorithms as referenced in Zhao et al. (2019, 2021).
- Copula-Based Outlier Detector (COPOD): Constructs a distribution function and identifies outliers at the extremes of the distribution Li et al. (2020).
- Empirical Cumulative Distribution Functions (ECOD): Uses a modeled step-function Li et al. (2022).
- Kernel Density Estimation for Unsupervised Outlier Detection (KDE): Employs the negative log probability density Latecki et al. (2007).
- Outlier detection based on Sampling: Uses statistical sampling techniques Sugiyama et al. (2013).
- Stochastic Outlier Selection (SOS): Uses the concept of affinity for proportional similarity between data points Janssens et al. (2012).
- Principal Component Analysis Outlier Detector (PCA): Employs a linear dimensionality reduction using singular value decomposition Shyu et al. (2003), Aggarwal et al. (2016).
- Linear Model Deviation-based outlier detection (LMDD): Uses a smoothing factor to indicate dissimilarity Arning et al. (1996), Zhao et al. (2019).
- One-class Support Vector Machine detector (OCSVM): Optimizes a high-dimensional distribution controlled through weight vectors Schölkopf et al. (2001).
- Nearest Neighbors Detector using the mean distance (AvgKNN): Uses the mean distance to neighbors as the outlier score Angiulli et al. (2002), Ramaswamy et al. (2000).
- Clustering Based Local Outlier Factor (CBLOF): Recasts detection as a clustering problem and calculates the distance to the nearest large cluster He et al. (2003).
- Connectivity-Based Outlier Factor (COF): Yields the ratio of average chaining distance for nearest neighbors Tang et al. (2002).
- Histogram-based Outlier Detection (HBOS): Uses binning according to the Birge-Rozenblac method Birgé et al. (2006), Goldstein et al. (2012).
- k-Nearest Neighbors Detector (KNN): Measures the distance between neighbors Angiulli et al. (2002), Ramaswamy et al. (2000).
- Local Outlier Factor (LOF): Measures the local deviation of density with respect to its neighbors Breunig et al. (2000).
- Nearest Neighbors Detector using the median distance (MedKNN): Uses the median distance to neighbors as the outlier score Angiulli et al. (2002), Ramaswamy et al. (2000).
- Subspace Outlier Detection (SOD): Compares axis-parallel subspaces in a high-dimensional feature space Kriegel et al. (2009).
- Feature bagging detector (FB): Combines multiple outlier detection methods and maximizes the scores Lazarevic et al. (2005).
- IsolationForest Outlier Detector (IForest): Analyzes path lengths in created tree structures of features Liu et al. (2008, 2012).
- Lightweight on-line detector of anomalies (LODA): Leverages an ensemble of weak detectors Pevny et al. (2016).
- Scalable Unsupervised Outlier Detection (SUOD): Optimizes a modular acceleration system Zhao et al. (2021).
- Anomaly Detection with Generative Adversarial Networks (AnoGAN): Where two artificial neural networks compete with each other to make accurate outlier predictions Schlegl et al. (2017).
- Fully-connected Auto Encoder (AE): For dimensionality reduction and outlier detection in latent space Aggarwal et al. (2016).
- Deep One-Class Classification for outlier detection (DeepSVDD): Trains an artificial neural network while minimizing the volume of a hyper-sphere that surrounds the data and calculating the distance to the center Ruff et al. (2018).
- Single-Objective Generative Adversarial Active Learning (SO-GAAL): Based on a mini-max game between generator and discriminator networks Liu et al. (2019).
- Variational Auto Encoder (VAE): For continuous representations in the latent space for reducing the dimensionality Kingma et al. (2013), Burgess et al. (2018).