CORSET: concise and interpretable multi-label rule sets

CORSET is an algorithm to learn a set of classification rules for multi-label classification problems.

The rules learned by CORSET are optimized for conciseness, interpretability, and (of course) predictive performance.

If you find the software or the paper useful, please consider citing our work:

@inproceedings{ciaperoni2022concise,
  title={Concise and interpretable multi-label rule sets},
  author={Ciaperoni, Martino and Xiao, Han and Gionis, Aristides},
  booktitle={IEEE International Conference on Data Mining},
  year={2022}
}

Environment setup

warning: tested only on macOS Monterey and Python 3.8.15

Requirement:

• some compiler for C/C++ programs, e.g., g++

Create the virtual environment and install the packages

python3 -mvenv .venv
source .venv/bin/activate 
pip install -r requirements.txt
make build

To make sure the setup is correct, run the tests:

pytest ./tests

An example: training and prediction

from corset.data import Dataset
from corset.greedy import GreedyCFTPDivMax
from corset.evaluation import Evaluator

# load the data
ds = Dataset(name='medical', datadir="./data", split_train=False)
ds.load()

# initialize the model
clf = GreedyCFTPDivMax(
    min_feature_proba=.8,
    min_label_proba=.8,
    n_tails_per_iter=100,
    n_heads_per_tail=10,
    lambd=10.0,
    n_max_rules=100
)

# fit it
clf.fit(ds.trn_X, ds.trn_Y)

# make predictions
pred_Y = clf.predict(ds.tst_X)

# evaluate the performance
evaluator = Evaluator()
perf = evaluator.report(pred_Y, ds.tst_Y)
print(perf)
# you should get something like:
# OrderedDict([('hamming_accuracy', 0.9889457523029682), ('subset_accuracy', 0.6484135107471852), ('micro_precision', 0.7616300036062027), ('micro_recall', 0.8716467189434586), ('micro_f1', 0.812933025404157), ('macro_precision', 0.8894680208607824), ('macro_recall', 0.5024357158149877), ('macro_f1', 0.4691918958698777)])

Faster heuristics

We provide a few alternatives to the original CORSET algorithm, aiming at scalability:

• corset.greedy.GreedyCFTPDivMaxV2: which only does the 1st round of greedy picking, instead of both rounds
• corset.greedy.GreedyCFTPDivMaxV3: an extension of corset.greedy.GreedyCFTPDivMaxV2 but sqrt is applied to the quality term.

Note that approximation guarantee is lost in these cases.

More examples?

You may check the following directories for more examples:

• ./notebooks
• ./scripts