(manuscript available upon request)
Final project for cs6362: advanced machine learning
Motivating question: How do Variational Causal Networks (VCNs) approximation of the posterior over Direct Acyclic Graphs (DAGs) in causal structure learning compare to the boostrapped estimation via Direct LiNGAM in high dimensional settings?
VCN adapted from: https://github.com/yannadani/vcn_pytorch (instructions located under models/vcn_adapted)
Boostrapped DirectLiNGAM adapted from: https://github.com/cdt15/lingam
Data (and verified ground truth source):
Line # | Intervention |
---|---|
0-852 | cd3_cd28 |
853-1754 | icam2 |
1755-2665 | aktinhib |
2666-3388 | g0076 |
3389-4198 | psitect |
4199-4997 | u0126 |
4998-5845 | ly |
5846-6758 | pma |
6759-7465 | b2camp |
Future work: ELBO gradient --> score function estimator variance reduction for more optimal learning within VCNs using control variates.