CounterFactual Regression with Importance Sampling Weights

Authors: Negar Hassanpour, Russell Greiner

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on two publicly available benchmarks demonstrate that the proposed method significantly outperforms state-of-the-art. 4 Experiments. Table 1: ENo RMSE, PEHE, and ϵATE performance measures (lower is better), each of the form mean (standard deviation) on the IHDP benchmark. Table 2: Aggregated ENo RMSE (lower is better) on the ACIC 18 benchmark.
Researcher Affiliation Academia Negar Hassanpour and Russell Greiner Department of Computing Science, University of Alberta, Canada {hassanpo, rgreiner}@ualberta.ca
Pseudocode Yes Algorithm 1 CFR-ISW: Counter Factual Regression with Importance Sampling Weights
Open Source Code No The paper mentions in the Acknowledgements: 'We wish to thank Dr. Martha White and Junfeng Wen for fruitful conversations, and Dr. Fredrik Johansson for publishing/maintaining the code-base for the CFR method online.' This refers to the code for the CFR method (a baseline), not the code for the authors' proposed CFR-ISW method.
Open Datasets Yes To make performance comparison easier, however, we do not synthesize our own datasets here. Instead, we use two publicly available benchmarks see Sec. 4.3. Infant Health and Development Program (IHDP) [...] We worked with the same dataset provided by and used in [Shalit et al., 2017; Johansson et al., 2016; Johansson et al., 2018]. Atlantic Causal Inference Conference 2018 (ACIC 18) [...] covariates matrix for each of these datasets are sub-sampled from a covariates table of real-world medical measurements taken from the Linked Birth and Infant Death Data (LBIDD) [Mac Dorman and Atkinson, 1998].
Dataset Splits Yes We report the methods performances by averaging over 100 realizations of outcomes with 63/27/10 train/validation/test splits.
Hardware Specification No The paper does not provide specific details about the hardware used for the experiments (e.g., CPU, GPU models, memory, or cloud instance types).
Software Dependencies No The paper mentions 'gradient descent optimizer', 'Adam optimizer', 'elu as the non-linear activation function', and 'Maximum Mean Discrepancy (MMD)' but does not provide specific version numbers for any software libraries, frameworks, or tools used (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes Hyperparameter Selection. Table 3: Hyperparameters and ranges (e.g., Imbalance parameter α {1E{-2, -1, 0, 1}}, Num. of representation layers {3, 5}, Batch size {100, 300}). We trained CFR-ISW s π0 logistic regression function with gradient descent optimizer and a learning rate of 1E-3. For both CFR and CFR-ISW, we trained the Φ and ht networks with regularization coefficient λ=1E-3, elu as the non-linear activation function, Adam optimizer [Kingma and Ba, 2015], learning rate of 1E-3, and maximum number of iterations of 3000.