Skip to content

Latest commit

 

History

History
43 lines (32 loc) · 4.1 KB

File metadata and controls

43 lines (32 loc) · 4.1 KB

Algorithms

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).

Probabilistic Detectors

  • 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).

Linear Models

Proximity-based Techniques

Ensembles

  • 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).

Neural Networks

  • 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).