We address the estimation of conditional average treatment effects (CATEs) for structured treatments (e.g., graphs, images, texts). Given a weak condition on the effect, we propose the generalized Robinson decomposition, which (i) isolates the causal estimand (reducing regularization bias), (ii) allows one to plug in arbitrary models for learning, and (iii) possesses a quasi-oracle convergence guarantee under mild assumptions. In experiments with small-world and molecular graphs we demonstrate that our approach outperforms prior work in CATE estimation.

Link to paper

We tested the implementation in Python 3.8.

requirements.txt is an automatically generated file with all dependencies.

Essential packages include:

rdkit
numpy
networkx
scikit-learn
torch
pyg
wandb

The TCGA simulation requires the TCGA and QM9 datasets. The code automatically downloads and unzips these datasets if they do not exist. Alternatively, the TCGA dataset can be downloaded from here and the QM9 dataset from here. Both datasets should be located in data/tcga/.

There are three runnable python scripts:

• generate_data.py: Generates and saves a dataset given the configuration in configs/generate_data/.
  • Stores generated data in data_path with folder structure {data_path}/{task}/seed-{seed}/bias-{bias}/
  • For each task, seed, and bias combination, generates and stores a new dataset
• run_model_training.py: Trains and evaluates a CATE estimation model given the configuration in configs/run_model/.
  • Evaluation results will be logged, can be saved to results_path and/or synced to a wandb.ai account
• run_hyperparameter_sweeping.py Sweeps hyper-parameters with wandb as specified in configs/sweeps/
• run_unseen_treatment_update.py: Runs the GNN baseline on a specified dataset and updates one-hot encodings of previously unseen treatments in the test set to the closest ones seen during training based on their Euclidean space in the hidden embedding space.
  • Before running the CAT baseline, run this script. Otherwise, unseen treatment one-hot encodings will be fed into the network.

• task: Simulation sw or tcga
• bias: Treatment selection bias coefficient
• seed: Random seed
• data_path: Path to save/load generated datasets

• task: Simulation sw or tcga
• model: SIN, gnn, cat, graphite, zero
• bias: Treatment selection bias coefficient
• seed: Random seed

When parsing smiles from the QM9 dataset for simulating a TCGA experiment, there may be bad input warnings for certain molecules. The data generator will ignore these molecules. When subsampling 10k molecules, we noticed that there are around ~1% faulty molecules.

For hyper-parameter tuning and experiment management, we use the wandb package. Please note that for both tasks, you need an account on wandb.ai. If you want to run single experiments, you can do so without an account - in this case, please ignore the warnings.