Learning Structural Causal Models from Ordering: Identifiable Flow Models
Authors: Minh Khoa Le, Kien Do, Truyen Tran
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we demonstrate that our method outperforms previous state-of-the-art approaches and delivers consistent performance across a wide range of structural causal models in answering observational, interventional, and counterfactual questions. Additionally, our method achieves a significant reduction in computational time compared to existing diffusion-based techniques, making it practical for large structural causal models. We empirically demonstrate that our method outperforms competing approaches across a wide range of synthetic and real datasets, excelling in estimating both the mean and overall shape of interventional and counterfactual distributions. The experiments confirm that our parallel architecture is not only scalable but also maintains high performance as the complexity and size of the datasets increase. |
| Researcher Affiliation | Academia | 1 Applied Artificial Intelligence Institute (A2I2), Deakin University, Australia EMAIL |
| Pseudocode | Yes | Algorithm 1: P-CFM Model Training. Input: Observational data X , causal ordering π, exogenous distribution p(u) 1: while not converged do Training S-CFM 2: z0 = x X; z1 = u p(u); t Unif[0, 1]; zt = (1 t) z0 + t z1 3: Update parameters of velocity neural network vθ, by minimizing the following loss: vθ (zt, x, t) (z1 z0) 2 4: end while 5: while not converged do Training Endogenous Predictor 6: t Unif[0, 1]; z0 = x X 7: Sample zt, z1 given z0 from S-CFM 8: Update parameters of endogenous predictor neural network NNθ, by minimizing the following loss: EPθ (zt, z1, t) z0 2 9: end while |
| Open Source Code | No | The paper does not provide explicit access to source code or a repository link. It only mentions 'We re-implement these methods based on provided code.' in reference to baselines. |
| Open Datasets | Yes | We evaluate S-CFM on real-world setting on the electrical stimulation interventional f MRI data (Thompson et al. 2020), experimental setup is obtained from (Khemakhem et al. 2021). ... Thompson, W. H.; Nair, R.; Oya, H.; Esteban, O.; Shine, J.; Petkov, C.; Poldrack, R.; Howard, M.; and Adolphs, R. 2020. A data resource from concurrent intracranial stimulation and functional MRI of the human brain. Scientific data, 7(1): 258. |
| Dataset Splits | Yes | We randomly sample 50,000/5,000/5,000 samples for training, evaluation and testing, respectively. |
| Hardware Specification | No | The paper mentions 'inference time (in seconds) against number of nodes d' and discusses 'computational time' but does not specify any particular CPU, GPU, or other hardware used for these computations. |
| Software Dependencies | No | The paper mentions using the 'Adam optimizer' and 'ELU activation' but does not provide specific version numbers for any software libraries, frameworks, or programming languages used (e.g., PyTorch, TensorFlow, Python version). |
| Experiment Setup | Yes | For MAVEN and endogenous predictor EPθ, we use a MADE with three hidden layers [256, 256, 256] and ELU activation, and a fully connected neural network with the same layers and activation. We use the Adam optimizer with a learning rate of 0.001 and apply a decay factor of 0.95 to the learning rate if it remains at a plateau for more than 10 epochs, batch size of 2048, and train for 900 epochs. Inference process is done in 50 steps. |