A Distillation Approach to Data Efficient Individual Treatment Effect Estimation
Authors: Maggie Makar, Adith Swaminathan, Emre Kıcıman4544-4551
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Using 77 semi-synthetic datasets with varying data generating processes, we show that DEITEE achieves significant reductions in the number of variables required at test time with little to no loss in accuracy. Using real data, we demonstrate the utility of our approach in helping soon-to-be mothers make planning and lifestyle decisions that will impact newborn health. |
| Researcher Affiliation | Collaboration | Maggie Makar CSAIL, MIT Cambridge, MA mmakar@mit.edu Adith Swaminathan Microsoft Research Redmond, WA adswamin@microsoft.com Emre Kıcıman Microsoft Research Redmond, WA emrek@microsoft.com |
| Pseudocode | No | The paper describes the algorithm steps in prose and mathematical formulations but does not provide a clearly labeled 'Algorithm' or 'Pseudocode' block. |
| Open Source Code | No | The paper does not contain any explicit statement about open-sourcing their code or a link to a repository for their implementation. |
| Open Datasets | Yes | For our semi-synthetic experiments, we use data generated for the Atlantic Causal Inference Conference Competition (Dorie et al. 2017). ... We use data from the 1989 Linked birth-infant death data which is made publicly available by the Centers for Disease Control (CDC ). The dataset has infant birth weight, as well as parent demographics and mother risk factors for all 4 million babies born in the US. ... https://www.cdc.gov/ nchs/nvss/linked-birth.htm. |
| Dataset Splits | Yes | We split the data into 2/3 for training and validation and 1/3 for testing. ... During training time, we do three-fold cross validation for each of the base models to pick the optimal hyper-parameters. |
| Hardware Specification | No | The paper does not mention any specific hardware used for running the experiments. |
| Software Dependencies | No | The paper mentions 'Details about hyper-parameter tuning and software used are in the appendix,' but no specific versioned software dependencies are listed in the main text of the paper. |
| Experiment Setup | Yes | We split the data into 2/3 for training and validation and 1/3 for testing. For each of the 77 simulated setups, we run DEITEE on one of three base estimates: oracle, Bayesian Additive Regression Trees (BART; Hill 2011), and Generalized Random Forests (GRF; Athey, Tibshirani and Wager 2016). ... During training time, we do three-fold cross validation for each of the base models to pick the optimal hyper-parameters. ... We then distill each base method as outlined previously using a decision tree, stopping the splitting when the improvement in accuracy is less than ϵ = 0.001. We run each experiment 20 times, each time randomly picking new simulation parameters and present results averaged over these 20 unique simulations. |