Variational Concept Embedding Model (V-CEM)

This repository contains the official implementation to reproduce the experiments in the paper:
V-CEM: Bridging Performance and Intervenability in Concept-based Models. Published at The 3rd World Conference on eXplainable Artificial Intelligence, 2025.

TL;DR

Concept-based eXplainable AI (C-XAI) improves model interpretability by using human-understandable concepts, enabling transparency and user intervention. Concept Bottleneck Models (CBMs) allow intervention by explicitly predicting concepts but underperform compared to black-box models. Concept Embedding Models (CEMs) improve in-distribution (ID) accuracy by learning from inputs and concept predictions but are less effective for intervention, especially out-of-distribution (OOD). We introduce the Variational Concept Embedding Model (V-CEM), which uses variational inference to boost intervention responsiveness. Evaluated across textual and visual datasets, V-CEM maintains CEM-level ID accuracy and matches CBM intervention performance in OOD, narrowing the gap between interpretability and generalization.

Reproducing the Experiments

To execute the code, please follow the detailed steps outlined below:

1. Create the virtual environment: First, you need to set up the virtual environment using the provided environment.yml file. This can be done by running the following command:

   conda env create -f environment.yml

2. Activate the environment: Once the virtual environment is created, you need to activate it. Use the command below to activate the environment named v_cem:

   conda activate v_cem

3. Run the experiments: With the environment activated, you can now run the experiments by executing the main script. Use the following command to start the experiments:

   python main.py

Citation

@article{de2025v,
   title={V-CEM: Bridging Performance and Intervenability in Concept-based Models},
   author={De Santis, Francesco and Ciravegna, Gabriele and Bich, Philippe and Giordano, Danilo and Cerquitelli, Tania},
   journal={arXiv preprint arXiv:2504.03978},
   year={2025}
}