Defining Expertise: Applications to Treatment Effect Estimation

Authors: Alihan Hüyük, Qiyao Wei, Alicia Curth, Mihaela van der Schaar

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we demonstrate that: (i) the type and the amount of expertise present in a dataset significantly influences the performance of different methods for treatment effect estimation (Sec. 4.1), and (ii) it may be possible to classify datasets according to what type of expertise they reflect and thereby identify what methods might be more or less suitable for a given dataset we propose a pipeline that does this (Sec. 4.2).
Researcher Affiliation Academia Alihan H uy uk , Qiyao Wei , Alicia Curth, Mihaela van der Schaar University of Cambridge
Pseudocode No The paper describes methods and processes in paragraph form and a flow diagram, but does not include structured pseudocode or algorithm blocks with numbered steps typically found in algorithm descriptions.
Open Source Code Yes Moreover, the code for reproducing our main experimental results can be found at https://github.com/Qiyao Wei/Expertise and https://github.com/vanderschaarlab/Expertise.
Open Datasets Yes Inspired by the simulator in Crabb e et al. (2022), and similar to them, we start with covariates X Rd from real-world datasets. [...] In the environments based on the TCGA dataset (Weinstein et al., 2013; Schwab et al., 2020)... Meanwhile, the News dataset (Newman, 2008)...
Dataset Splits Yes All models are trained using the Adam optimizer with learning rate 0.001, batch size 1024, and early stopping on a validation set, where we employ a standard train-validation split of 70% 30%.
Hardware Specification Yes We used a virtual machine with six 6-Core Intel Xeon E5-2690 v4 CPUs, one Tesla V100, and 110GB of RAM to run all experiments.
Software Dependencies No The paper mentions "Py Torch implementations" and the "python package CATENets" but does not specify their version numbers or other software dependencies with explicit version details.
Experiment Setup Yes All models are trained using the Adam optimizer with learning rate 0.001, batch size 1024, and early stopping on a validation set, where we employ a standard train-validation split of 70% 30%.