Optimal Transport for Treatment Effect Estimation
Authors: Hao Wang, Jiajun Fan, Zhichao Chen, Haoxuan Li, Weiming Liu, Tianqiao Liu, Quanyu Dai, Yichao Wang, Zhenhua Dong, Ruiming Tang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments |
| Researcher Affiliation | Collaboration | 1Zhejiang University 2Tsinghua University 3Peking University 4Purdue University 5 Huawei Noah s Ark Lab |
| Pseudocode | Yes | The optimization procedure consists of three steps as summarized in Algorithm 3 (see Appendix B). First, compute πϵ,κ,γ by solving the OT problem in Definition 3.2 with Algorithm 2 (see Appendix B)... |
| Open Source Code | No | The paper does not contain an explicit statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | IHDP and ACIC. Specifically, IHDP is designed to estimate the effect of specialist home visits on infants potential cognitive scores, with 747 observations and 25 covariates. ACIC comes from the collaborative perinatal project [51], and includes 4802 observations and 58 covariates. |
| Dataset Splits | No | The paper mentions 'within-sample and out-of-sample results' on 'training and test set' and 'validate performance every two epochs', implying the use of validation, but it does not specify the exact percentages or counts for training, validation, or test splits, nor does it provide citations to specific predefined splits that would define these ratios. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments. |
| Software Dependencies | No | The paper mentions 'Adam optimizer' and refers to Kingma and Ba [35] for 'Other settings of optimizers' but does not specify any software dependencies with version numbers. |
| Experiment Setup | Yes | A fully connected neural network with two 60-dimensional hidden layers is selected to instantiate the representation mapping ψ and the factual outcome mapping ϕ for ESCFR and other neural network based baselines. To ensure a fair comparison, all neural models are trained for a maximum of 400 epochs using the Adam optimizer, with the patience of early stopping being 30. The learning rate and weight decay are set to 1e 3 and 1e 4, respectively. |