Learning Triggers for Heterogeneous Treatment Effects

Authors: Christopher Tran, Elena Zheleva5183-5190

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results on multiple datasets show that our approach can learn the triggers better than existing approaches.
Researcher Affiliation Academia Christopher Tran, Elena Zheleva Department of Computer Science, University of Illinois at Chicago Chicago, IL, USA {ctran29, ezheleva}@uic.edu
Pseudocode Yes Algorithm 1 Learning trigger-based causal trees
Open Source Code Yes Code available: https://github.com/chris-tran-16/CTL
Open Datasets Yes We study two datasets that lend themselves to the trigger-based treatment problem which is the focus of our work. We also use the ACIC Causal Inference Challenge dataset for binary treatments. For the first dataset, we use data from the 1987 National Medical Expenditure Survey (NMES) (Johnson et al. 2003). The second dataset is a cloud storage usage survey, which asks users various questions about their cloud storage (Khan et al. 2018). The last dataset comes from the ACIC 2018 Causal Inference Challenge2.
Dataset Splits Yes For each dataset we do an 80-20 randomized split for the training and testing set for evaluation. For ours and honest trees, we split the training into two parts: a training Str and validation Sval. For our approach, we clearly separate a training, validation, and testing sample for training and evaluation. We build the tree on the training portion, while penalizing generalization ability based on the validation set. We learn λ and the validation split size for our methods on a separate validation set.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU, GPU models, memory, or specific computing platforms) used to run the experiments.
Software Dependencies No The paper does not list specific versions of software libraries, frameworks, or programming languages used (e.g., Python version, specific machine learning library versions).
Experiment Setup Yes For each dataset we do an 80-20 randomized split for the training and testing set for evaluation. For ours and honest trees, we split the training into two parts: a training Str and validation Sval. We grow a maximum sized tree and prune all trees based on statistical significance (α = 0.05). We learn λ and the validation split size for our methods on a separate validation set.