Causal Inference for Event Pairs in Multivariate Point Processes

Authors: Tian Gao, Dharmashankar Subramanian, Debarun Bhattacharjya, Xiao Shou, Nicholas Mattei, Kristin P Bennett

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct an experimental investigation using synthetic and real-world event datasets, where our proposed causal inference framework is shown to exhibit superior performance against a set of baseline pairwise causal association scores.
Researcher Affiliation Collaboration Tian Gao IBM Research tgao@us.ibm.com Dharmashankar Subramanian IBM Research dharmash@us.ibm.com Debarun Bhattacharjya IBM Research debarunb@us.ibm.com RPI shoux@rpi.edu Nicholas Mattei Tulane University nsmattei@tulane.edu Kristin Bennett RPI bennek@rpi.edu
Pseudocode Yes Algorithm 1 Inverse Probability Weighting for Events
Open Source Code No The code will be released in Github.
Open Datasets Yes We begin by comparing the ATE estimation performance of our proposed IPTW methods on synthetic event datasets, generated using different parameters. [...] We also test our methods on the diabetes dataset [14] a real-world dataset which we process into events for meals, exercise activity, insulin dosage and changes in blood glucose measurements for 70 diabetes patients.
Dataset Splits Yes The dataset is split into 50%/50% training/test sets, and optimal window setting is determined on the training set, which is then deployed in the test set for evaluation.
Hardware Specification Yes All experiments are done on a machine with 2.9 GHz quad-core CPU.
Software Dependencies No The paper mentions using 'tick: A Python library' [6] and 'PGEM' [7] but does not specify version numbers for these or other software dependencies.
Experiment Setup Yes We generate 3 models with different numbers of events, randomly generated graph structures among events, fixed window size of w = 30, T = 2000, and random intensities between 0.1 and 0.4. We use the data and the generated model to obtain the true estimates of λy|Zt(t) at chosen times t and hence can compute the ground truth ATE. Since we observed that the sample size S of t (103 to 105) in the ATE estimation does not impact the results much, we use sample size S = 103 for all our experiments.