bidag_order_mcmc
Order MCMC
Package |
|
Version |
2.1.4 |
Language |
|
Docs |
|
Paper |
|
Graph type |
|
Docker |
|
Module folder |
Description
This technique relies on a Bayesian perspective on structure learning, where the score of a DAG is defined as its posterior distribution. To overcome the limitation of simple structure-based MCMC schemes, Friedman and Koller[1] turned to a score defined as the sum of the posterior scores of all DAG which are consistent with a given topological ordering of the nodes. One can then run a Metropolis-Hasting algorithm to sample from the distribution induced by the order score, and later draw a DAG consistent with the order. This strategy substantially improves convergence with respect to earlier structure MCMC scheme, though it unfortunately produces a biased sample on the space of DAGs. The implementation considered in Benchpress is a hybrid version with the sampling performed on a restricted search space initialised with constraint-based testing and improved with a score-based search Kuipers et al.[2].
Some fields described
input_algorithm_id
Algorithm to use for initial search space. This should be the ID of another algorithm object. It corresponds to the original startspace parameter in the R package.
Example JSON
[
{
"id": "omcmc_itmap-bge",
"input_algorithm_id": "itsearch_map-bge_am01_endspace",
"plus1": true,
"scoretype": "bge",
"chi": null,
"edgepf": null,
"aw": null,
"am": [
0.01,
0.1,
0.05
],
"alpha": 0.01,
"gamma": 1,
"stepsave": null,
"iterations": null,
"MAP": true,
"cpdag": false,
"timeout": null,
"mcmc_seed": 1,
"threshold": 0.5,
"mcmc_estimator": "threshold",
"burnin_frac": 0.5
},
{
"id": "omcmc_itmap-bde",
"plus1": true,
"input_algorithm": "itsearch_map-bde",
"scoretype": "bde",
"chi": [
0.01,
0.1,
1,
2
],
"edgepf": 2,
"aw": null,
"am": null,
"alpha": 0.05,
"gamma": 1,
"stepsave": null,
"iterations": null,
"MAP": true,
"cpdag": false,
"mcmc_seed": 1,
"threshold": [
0.5
],
"burnin_frac": 0,
"mcmc_estimator": "threshold",
"timeout": null
}
]