dabrze / measure_dynamics

Code and scripts accompanying the paper: "On the Dynamics of Classification Measures for Imbalanced Data"

Measure Dynamics

Source codes, reproducible experiment scripts, and numerical results accompanying the paper: "On the Dynamics of Classification Measures for Imbalanced Data" by Brzezinski, D. et al.

The repository is divided into three main sections (folders):

• matlab
• python
• java

matlab

The matlab folder contains code responsible for generating all possible confusion matrices for a given size of data set n, and visualizing this data in the form of histograms of measure values and gradients in barycentric space.

To run the code, simply run the file run_visualizations.m in matlab. After a certain set of visualizations is displayed, press Enter to generate the next set of images.

python

The python folder contains source code and data sets needed to run the measure normalization experiments. The following libraries are required:

• pandas
• numpy
• scipy
• scikit-learn

Package versions are listed in the requirements.txt file.

To run the code, use the following command:

python normalization_experiment.py

java

The java folder contains classes required to reproduce the drift detection and class ratio change experiments. The experiments run MOA.

To re-run the experiments include into MOA the classes from the moa folder and run java with the follwoing to classes: moa.experiments.PHTests.java and moa.experiments.RatioChangeTests.java. The plots for the ratio change experiment can be reproduced using gnuplot and the visualize.py script.