Meta-learning for heterogeneous treatment effect estimation with closed-form solvers

Paper

Codes

Usage

• Generate datasets by setting (ii) in Curth and van der Schaar (2021).
  % python crt_data_curth.py 10000 100 1
• Estimate pseudo-CATEs.
  % python estimate_tau.py --fn data/synth_curth_s10000t100I1/i0.npz --out data/synth_curth_s10000t100I1/i0tau
• Run meta-learning.
  % python train_meta_cate.py --fn data/synth_curth_s10000t100I1/i0tau.npz --method 0 --second_method 2 --n_spt0 3 --n_spt1 3 --GP 0 --epoch 5000
  options:
  • method: loss function for task-shared parameters (0:CATE loss, 1:regerssion loss)
  • second_method: base lernaer (0:RA, 1:PA, 2:DR, 3:plugin)
  • n_spt0: number of untreated instances in a support set
  • n_spt1: number of treated instances in a support set
  • GP: GP or linear layer for task-specific last layers (0:Linear, 1:GP)
  • epoch: number of epochs

License