Exploiting Geometry for Treatment Effect Estimation via Optimal Transport
Authors: Yuguang Yan, Zeqin Yang, Weilin Chen, Ruichu Cai, Zhifeng Hao, Michael Kwok-Po Ng
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental studies on both synthetic and real-world datasets demonstrate the effectiveness of our proposed method. We conduct extensive experiments on both synthetic and real-world data sets to demonstrate the advantages of our proposed method in terms of treatment effect estimation and robustness to outliers. |
| Researcher Affiliation | Academia | Yuguang Yan1, Zeqin Yang1, Weilin Chen1, Ruichu Cai1,2*, Zhifeng Hao3, Michael Kwok-Po Ng4 1School of Computer Science, Guangdong University of Technology, Guangzhou, China 2Peng Cheng Laboratory, Shenzhen, China 3College of Science, Shantou University, Shantou, China 4Department of Mathematics, Hong Kong Baptist University, Hong Kong, China |
| Pseudocode | Yes | Algorithm 1: Optimal Transport for Causal Inference. |
| Open Source Code | No | The paper does not provide an explicit statement or a link to an open-source code repository for the methodology described in the paper. |
| Open Datasets | Yes | La Londe1 consists of two parts. The first part comes from a RCT (NSW). In the second part, as (Kuang et al. 2017) did, we replace the control group in NSW with another control group from the observational data (CPS). The treatment is whether the participant attend the job training program, and the outcome is the earning in 1978. The data contains 8 covariates. Twins is collected from the twins born in USA between 1989-1991 (Almond, Chay, and Lee 2005). Each twin pair has 30 covariates. 1https://users.nber.org/~rdehejia/data/.nswdata2.html |
| Dataset Splits | No | The paper mentions generating samples for simulation data and conducting experiments multiple times, but it does not provide explicit training/validation/test dataset splits (e.g., percentages or counts) or details about cross-validation procedures. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used to run the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper does not list specific software dependencies with version numbers (e.g., 'Python 3.8, PyTorch 1.9') that would be necessary for reproducibility. |
| Experiment Setup | Yes | We take La Londe as an example to evaluate the parameter sensitivity of OTCI. We vary the parameters α and γc in Eq. (16), and plot the results in Fig. 3. From Fig. 3a, MAE increases when α becomes small or large, indicating that both inter and intra-group geometric information play important roles in removing confounding bias. From Fig. 3b, MAE increases when γc is larger than 10 2, since a large γc will push the learned weights close to the uniform distribution, resulting in a failure of confounding bias elimination. Overall, when parameters 0.4 α 0.7 and γc 1 10 2, OTCI achieves promising performance, which demonstrates its stability to the trade-off parameters in a certain range. |