Exogenous Matching: Learning Good Proposals for Tractable Counterfactual Estimation
Authors: Yikang Chen, Dehui du, Lili Tian
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments. We validate the theoretical results through experiments under various types and settings of Structural Causal Models (SCMs) and demonstrate the outperformance on counterfactual estimation tasks compared to other existing importance sampling methods. |
| Researcher Affiliation | Academia | Yikang Chen1, Dehui Du1,*, Lili Tian1 1Shanghai Key Laboratory of Trustworthy Computing, East China Normal University |
| Pseudocode | Yes | Algorithm 1 Exogenous Matching Learning. Algorithm 2 Multiple Proposals Inference. Algorithm 3 Finding Counterfactual Markov Boundary. |
| Open Source Code | Yes | Code is available at: https://github.com/cyisk/exom |
| Open Datasets | Yes | Markovian diffeomorphic SCMs These SCMs are sourced from [42] and are primarily used in the experiments for Causal NF, including: CHAIN-LIN-3, CHAIN-NLIN-3, CHAIN-LIN-4, CHAIN-LIN-5, COLLIDER-LIN, FORK-LIN, FORK-NLIN, LARGEBD-NLIN, SIMPSON-NLIN, SIMPSON-SYMPROD, TRIANGLE-LIN, and TRIANGLE-NLIN. ... Semi-Markovian continuous SCMs These SCMs are described in [107]. ... Regional canonical SCMs These SCMs are originated from [108]... |
| Dataset Splits | Yes | In our experiments, we selected a batch size of 256 and generated a training set of size 16,384 through the sampler. The validation set is also generated using this sampler. |
| Hardware Specification | Yes | All models, including the pre-trained SCM agents, were trained and tested on an NVIDIA RTX 4090 and Intel(R) Xeon(R) Gold 6430 platform. |
| Software Dependencies | No | The paper mentions software like 'Zuko library' and 'Adam W' but does not specify their version numbers or other crucial software dependencies with versions. |
| Experiment Setup | Yes | Specifically, we utilize the same sampler to construct PY , from which random y is sampled. For discrete distributions, counterfactual events are assigned as Y {y }, and for continuous distributions, counterfactual events are assigned as Y δl(y ), where δl(y ) is a cube centered at y with side length l. ... the number of components in GMM is fixed at 10, the number of transformations for all flow models is set to 5, and all the neural networks involved contain 2 hidden layers with 64 neurons each (in some experiments, 256 neurons were used, which will be specifically indicated and discussed in subsequent sections). ... we used Adam W [62] with an initial learning rate of 0.001 as the optimizer and Reduce On Plateau with a patience of 5 and a factor of 0.5 as the learning rate scheduler. |