Introduction

This algorithm is developed as part of my master's thesis: Anomaly Detection Using an Ensemble with Simple Sub-models, 2024. The algorithm explores the effectiveness of an ensemble of simple sub-models like linear regression in detecting anomalies.

Results from the thesis experiments

The AUROCs of the runs reported in the thesis are stored in this Google Sheet

Benchmarking

Installation

Install the package using pip.

pip install adess==1.0.0

Parameters

Run with --help to see the parameters

adess --help

usage: adess [-h] --train TRAIN --test TEST [--feat_sel_percent FEAT_SEL_PERCENT] [--max_feats MAX_FEATS] [--order ORDER] [--computation_budget COMPUTATION_BUDGET] [--no_submodels NO_SUBMODELS] [--prep PREP] [--extract EXTRACT]
             [--submodel SUBMODEL]

ADESS: Anomaly Detection using Ensemble of simple sub=models

options:
  -h, --help            show this help message and exit
  --train TRAIN         Training data in .npy file (default: None)
  --test TEST           Testing data in .npy file (default: None)
  --feat_sel_percent FEAT_SEL_PERCENT
                        Feature selection percentage (default: 0.2)
  --max_feats MAX_FEATS
                        Maximum number of features (default: 50)
  --order ORDER         Degree of polynomials for feature bagging (default: 2)
  --computation_budget COMPUTATION_BUDGET
                        Computation budget in seconds (default: 600)
  --no_submodels NO_SUBMODELS
                        Count of submodels in the ensemble (default: 500)
  --prep PREP           List of preprocessing options (choose one or many): [skel,canny,clahe,blur,augment,gray,norm,std,none] (default: ['norm'])
  --extract EXTRACT     Feature selection option (choose one): [rbm,tsne,pca,ica,nmf,ae,none] (default: pca)
  --submodel SUBMODEL   Submodel type option (choose one): [lin,lasso,elastic,svm] (default: lin)

Usage: CLI

adess --train path/to/train --test path/to/test

Example:

adess --train train.npy --test test.npy

Output:

X_train.shape = (353, 10), X_test.shape = (89, 10), 'feat_sel_percent = 0.2', 'max_feats = 50', 'order = 2', 'computation_budget = 600', 'no_submodels = 500', 'prep = norm', 'extract = pca', 'submodel = lin'
100%|█████████████████████████████████████████| 500/500 [00:00<00:00, 1094.41it/s]
Mean of Predicted Y = 8.005149077019986e-32, Count of submodel executed = 500

Usage: Python

1. Import the sklearn diabetes dataset as an example.
2. Split and load the dataset to the adess() function. X_test will be used to predict 'y'.
3. The mean prediction (of y) and the (default) ensemble size are printed.

>>> from adess.adess import adess
>>> from sklearn.datasets import load_diabetes
>>> from sklearn.model_selection import train_test_split
>>> X, y= load_diabetes(return_X_y=True)
>>> X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)
>>> adess(X_train,X_test)
100%|█████████████████████████████████████████| 500/500 [00:00<00:00, 1162.27it/s]
(3.2565381430216842e-31, 500)
>>>