DiffVersify: a scalable approach to differentiable pattern mining though coverage regularization

This is the python implementation accompanying the paper DiffVersify: a scalable approach to differentiable pattern mining though coverage regularization. The code is based on Fischer et al [1] and is available under the license GNU General Public License v3.0. We strongly recommend to run the code on a GPU.

Presentation

Pattern mining addresses the challenge of automatically identifying interpretable and discriminative patterns within data. Recent approaches, leveraging differentiable approach through neural autoencoder with class recovery, have achieved encouraging results but tend to fall short as the magnitude of the noise and the number of underlying features increase in the data. Empirically, one can observe that the number of discovered patterns tend to be limited in theses challenging contexts. In this article, we present a differentiable binary model which incorporates a novel regularization technique that emphasizes pattern coverage. Moreover, we introduce an innovative pattern decoding strategy taking advantage of non-negative matrix factorization (NMF), extending beyond conventional thresholding methods prevalent in existing approaches. Experimentally, across four real-world datasets, our method showcases superior performance in terms of the ROC-AUC metric. Then, we demonstrate its capability to enhance pattern detection by augmenting the similarity with the ground truth patterns for synthetic data. Finally, we propose several evaluation metrics to better characterize the appropriateness of discovered patterns relative to the data, with a focus on pattern coverage. Our results indicate the efficacy of our approach in this regard, demonstrating strong performance across various datasets.

Technical informations

Required packages

• pyTorch (1.12.1 for the results in the paper)
• numpy
• pandas
• scipy
• torchmetrics
• scikit-learn
• matplotlib
• seaborn
• tqdm
• tabulate
• openpyxl
• tensorboard

Folder organization

• code: Contains the code of our method, diffversify, and scripts to reproduce all the reults presented in the paper
• data: Data used for the experiments for real data.
• results: Directory to store the results outputted by the scripts in code
• appendix.pdf: Contains appendix for the main paper

Running experiments locally

To run the experiments on synthetic data, the scripts run_exp.pyallow to run all 4 synthetic experimennt and the real data one. The parameter --xp {xp} define which experimennt to run :
1 :exp1 - scalability over the number of features
2 :exp2 - scalability over the number of classes
3 :exp3 - robustness to additive noise
4 :exp4 - robustness to destructive noise
5 :expR - real dataset

For the xp 5 (real data) a second argument is necessary --dataset to specify over which dataset to run the real dataset experimennt. The available values are [cardio, disease, braca-n, brca-s].

ex : $python3 real_exp.py --xp 5 -d cardio ,

executes diffversify for the specified dataset. The results are saved to results/real_results/.

An optional --output [FOLDER_PATH] is available to compress and copy the results to a specified directory

Running experiment over docker

A dockerfile is joined to run the experiment remotely or without installing dependency. It is recommended to use the --output argument with a mounted folder to copy the results outside of the docker.
A public build docker is available at tchataing/diffversify

[1] Fischer, J.; and Vreeken, J. 2021. Differentiable pattern set mining. In Proceedings of the ACM International Conference on Knowledge Discovery and Data Mining (SIGKDD)