Continuous Treatment Effect Estimation Using Gradient Interpolation and Kernel Smoothing

Authors: Lokesh Nagalapatti, Akshay Iyer, Abir De, Sunita Sarawagi

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method on five benchmarks and show that our method outperforms six state-of-the-art methods on the counterfactual estimation error. 5 Experiments Dataset Following the prior work (Nie et al. 2021; Zhang et al. 2022; Bica, Jordon, and van der Schaar 2020), we use five datasets, namely IHDP, NEWS, and TCGA(0-2) on three types of treatments. Across all datasets, t [0, 1]. Details are deferred to the Appendix.
Researcher Affiliation Academia Indian Institute of Technology Bombay nlokesh@cse.iitb.ac.in, akshaygiyer@gmail.com, abir@cse.iitb.ac.in, sunita@iitb.ac.in
Pseudocode Yes Algorithm 1: GIKS training
Open Source Code Yes We release the code at: https://github.com/nlokeshiisc/GIKS release.
Open Datasets Yes We use five datasets, namely IHDP, NEWS, and TCGA(0-2) on three types of treatments. Details are deferred to the Appendix. Following the prior work (Nie et al. 2021; Zhang et al. 2022; Bica, Jordon, and van der Schaar 2020)
Dataset Splits Yes We allocate 30% samples as validation dataset to tune hyperparameters.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as GPU models, CPU types, or memory specifications.
Software Dependencies No The paper mentions using 'Adam W optimizer' and 'cosine kernel' but does not specify version numbers for any software dependencies or libraries (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes GIKS has three hyperparameters: learning rate, λGI, λGP that are optimized via grid search on factual error of the validation dataset. Further, the GP employs a cosine kernel. We use a batch size of 128, the Adam W optimizer, and early stopping based on factual error on the validation dataset. The results of hyperparameter tuning are presented in Table 3. TCGA(0-2) IHDP NEWS lrn rate 10 4 10 2 10 3 λGI 10 1 10 4 10 2 λKS 10 2 10 1 10 4 Table 3: The hyperparameters.