Analyze Classification Models by highlighting regions where performance is abnormal

Dependencies are managed with uv. Since it generates a requirements.txt, it should be compatible with other tools, like pip.

uv pip install -r requirements.txt

There are 5 parameters, 3 of which are required:

• A CSV containing the dataset. (required)
• A CSV containing the model's errors. Each instance should have an error associated with each class. You can use a binary approach: if the prediction for a given instance is right, all of it's errors are zero. Otherwise, if the prediction is wrong, the all of it's errors are zero except for the actual class, which receives an error of one. (required)
• The dataset's column that represents the output. By default, it's "target".
• The maximum number of subgroups to generate. By default, it's 20.
• A class from the dataset, to filter the subgroups for that specific class. (required)

This repo contains a sample dataset (iris) and a model's errors to provide an example execution:

source .venv/bin/activate # assuming your virtual env is located at .venv
python -m src --data ./data/iris.csv --errors ./data/errors/iris.csv --class setosa