/Unsupervised_Concept_Drift_Detectors_Analysis

Simulation, testing and comparison of state of the art Unsupervised Concept Drift Detectors used in a batch Machine Learning scenario.

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

Unsupervised Concept Drift Detectors Analysis

This project was developed during my master thesis at Politecnico di Milano, in collaboration with Bip.

Goal

The goal of the thesis is understanding the abilities of state of the arts unsupervised concept drift detection used in the scenario of the classic batch machine learning. To do so a testing framework was developed in order to simulate the aforementioned use case and the techniques found has been adapted to the working assumptions.

Unsupervised Concept Drift Detectors

The tested methodologies can be found here, they are:

Usage

All the experiments can be launched using main.py, inserting which models and data sets must be used and how many repetitions of the experiment perform.

The concept drift detectors will be launched with default hyper parameters, which can be modified in the implementation file of each methods (<method_name>.py) changing the values inside the "<method_name>_inference" function.

The results will be output in the 'result' folder as a json file for each data set.