Models Guide¶
List of Benchmarked Models¶
As of today, available models included in this benchmarking study are
Statistical models
Standard Deviation
Median Absolute Deviation (MAD)
Interquartile Range (IQR)
Z-score
Modified Z-score
Distance-based models
Euclidean Distance
Manhattan Distance
Minkowski Distance
Mahalanobis Distance
Machine-learning models
Isolation Forest
K-Nearest Neighbors (KNN)
Gaussian Mixture Models (GMM)
Local Outlier Factor (LOF)
Principal Component Analysis (PCA)
Autoencoders (AE)
Hyperparameter Tuning¶
In this study, we implemented hyperparameter tuning for unsupervised anomaly detection models using two different methods:
Leveraging meta-learning with labelled outliers to train a model with hyperparameter tuning (training dataset) and subsequenty predict the model performance on new, unlabeled datasets (test dataset) and
Using the output from the unsupervised model (e.g., predicted inliers from all anomaly detection models) as features for a supervised task, and adjust the unsupervised model’s hyperparameters to maximize the performance of the downstream supervised model.
Instead of blindly trying random hyperparameters as in random search or computing expensive hyperparameter searching as in grid search, we implemented Bayesian optimization to build a probabilistic model of the objective function and to choose the most promising hyperparameters.
Example¶
List of Model Tutorials