This repository contains the official implementation of LRI as described in the paper: Interpretable Geometric Deep Learning via Learnable Randomness Injection by Siqi Miao, Yunan Luo, Mia Liu, and Pan Li.

• Jan. 21, 2023: This paper is accepted to ICLR 2023!
• Oct. 20, 2022: This paper will show up at NeurIPS 2022 AI for Science Workshop!

This work systematically studies interpretable Geometric Deep Learning (GDL) models by proposing a framework Learnable Randomness Injection (LRI) and four datasets with ground-truth interpretation labels from real-world scientific applications in Particle Physics and Biochemistry.

We study the interpretability in GDL from the perspectives of existence importance and location importance of points, and instantiated LRI with LRI-Bernoulli and LRI-Gaussian to test the two types of importance, respectively. Fig. 1 shows the architectures of LRI.

The intuition is that if the existence of some points is important, then imposing large Bernoulli randomness on their existence will greatly affect the prediction loss; while if the geometric locations of some points are important, imposing large Gaussian randomness on their coordinates should also affect the prediction loss significantly. Therefore, to achieve great prediction accuracy, LRI will denoise those important points, and thus the learned randomness level measures the importance of the points.

Figure 1. The architectures of LRI-Bernoulli (top) and LRI-Gaussian (bottom).

All our datasets can be downloaded and processed automatically by running the scripts in ./src/datasets. By default, the code will ask if the raw files and/or the processed files should be downloaded. For example, to download and process the SynMol dataset, simply run:

cd ./src/datasets
python synmol.py

All datasets are also available to download from Zenodo manually: https://doi.org/10.5281/zenodo.7265547. Fig. 2 provides the illustrations of the four datasets, and Tab. 1 shows the statistics of them. We will update the description of each dataset with more details in README.md soon.

Figure 2. Illustrations of the four scientific datasets in this work to study interpretable GDL models.

Table 1. Statistics of the four datasets. $\mathbf{X}$ denotes point features, and $\mathbf{r}$ denotes geometric coordinates.

We have tested our code on Python 3.9 with PyTorch 1.12.1, PyG 2.0.4 and CUDA 11.3. Please follow the following steps to create a virtual environment and install the required packages.

Step 1: Clone the repository

git clone https://github.com/Graph-COM/LRI.git
cd LRI

Step 2: Create a virtual environment

conda create --name lri python=3.9 -y
conda activate lri

Step 3: Install dependencies

conda install -y pytorch==1.12.1 torchvision cudatoolkit=11.3 -c pytorch
pip install torch-scatter==2.0.9 torch-sparse==0.6.14 torch-cluster==1.6.0 torch-geometric==2.0.4 -f https://data.pyg.org/whl/torch-1.12.0+cu113.html
pip install -r requirements.txt

Use the following command to train a model:

cd ./src
python trainer.py --backbone [backbone_model] --dataset [dataset_name], --method [method_name]

backbone_model can be chosen from dgcnn, pointtrans and egnn.

dataset_name can be chosen from actstract_2T, tau3mu, synmol and plbind, and the dataset specified will be downloaded automatically.

method_name can be chosen from lri_bern, lri_gaussian, gradcam, gradgeo, bernmask, bernmask_p, and pointmask.

By adding --cuda [GPU_id] to the command, the code will run on the specified GPU; by adding --seed [seed_number] to the command, the code will run with the specified random seed.

The tuned hyperparameters for all backbone models and interpretation methods can be found in ./src/config.

If you find our paper and repo useful, please cite our paper:

@article{miao2023interpretable,
  title       = {Interpretable Geometric Deep Learning via Learnable Randomness Injection},
  author      = {Miao, Siqi and Luo, Yunan and Liu, Mia and Li, Pan},
  journal     = {International Conference on Learning Representations},
  year        = {2023}
}