Learnable Concept-Based Model

This repository contains the official implementation to reproduce the experiments in the paper: Towards Better Generalization and Interpretability in Unsupervised Concept-Based Models. Published at ECML PKDD 2025 (Research Track).

---

Paper Summary

We introduce LCBM (Learnable Concept-Based Model), a novel unsupervised concept-based model for image classification that improves both accuracy and interpretability.

Unlike prior approaches, LCBM learns a compact set of concepts represented as Bernoulli latent variables, each associated with an embedding to overcome the trade-off between performance and interpretability. LCBM achieves strong performance across multiple datasets and supports interpretable predictions through linear combinations of concept activations.

Key Improvements

• Improved generalization: Matches or exceeds prior unsupervised CBMs and approaches black-box performance.
• Human-aligned concepts: Concepts are more intuitive and better aligned, as shown through F1 scores, CAS, and a user study.
• Faithful explanations: Supports concept interventions and visual dictionaries for transparent decision-making.

We evaluated the model on both toy and real datasets: MNIST Even/Odd, MNIST Addition, CIFAR-10, CIFAR100, Tiny ImageNet, Skin Lesions and CUB-200.

Reproducibility Instructions

Follow the steps below to reproduce the results reported in the paper.

1. Environment Setup

Make sure you have conda installed.

conda env create -f environment.yaml
conda activate lcbm

2. Training Models

To train and save all models used in the paper:

• LCBM (ours):

bash bash_scripts/run_LCBM.sh

Baselines:

• BlackBox:

bash bash_scripts/run_E2E.sh

• Label-Free CBM:

bash bash_scripts/run_LF_CBM.sh

For this baseline, the concepts have been extracted from GPT-4o for each dataset and are reported into the concept_list folder.

⚠️ For other baselines (SENN, BotCL, etc.), we used the official repositories. We recommend doing the same for faithful reproduction.

3. Evaluation & Explanations

To compute all the metrics and generate explanations:

bash bash_scripts/compute_concept_metrics.sh
bash bash_scripts/run_explanations.sh

These scripts cover concept alignment, interpretability metrics, and qualitative visualizations such as concept dictionaries and Grad-CAMs.

Citation

@article{de2025towards,
  title={Towards Better Generalization and Interpretability in Unsupervised Concept-Based Models},
  author={De Santis, Francesco and Bich, Philippe and Ciravegna, Gabriele and Barbiero, Pietro and Giordano, Danilo and Cerquitelli, Tania},
  journal={arXiv preprint arXiv:2506.02092},
  year={2025}
}