causalcompass.algorithms.CMLP

class causalcompass.algorithms.CMLP(lag=3, hidden_dim=[100], lam=0.005, lr=0.01, max_iter=50000, lam_ridge=1e-2, penalty='H', device='cuda', seed=None)[source]

Deep learning-based causal discovery method that uses component-wise MLPs to model nonlinear Granger causality.

References

https://github.com/iancovert/Neural-GC

Parameters:

  • lag (int, default 3) – Maximum time lag

  • hidden_dim (list, default [100]) – Number of hidden units per layer

  • lam (float, default 0.005) – Sparsity penalty term parameter

  • lr (float, default 0.01) – Learning rate

  • max_iter (int, default 50000) – Maximum training iterations

  • lam_ridge (float, default 1e-2) – Ridge regularization parameter

  • penalty (str, default 'H') – Penalty type: ‘GL’ (group lasso), ‘GSGL’ (group sparse group lasso), or ‘H’ (hierarchical)

  • device (str, default 'cuda') – Computation device

Examples

>>> from causalcompass.algorithms import CMLP
>>> model = CMLP(lag=3, hidden_dim=[100], lam=0.005, lr=0.01, max_iter=50000, device='cuda')
>>> predicted_adj = model.run(X)
>>> all_metrics, no_diag_metrics = model.eval(true_adj, predicted_adj)
__init__(lag=3, hidden_dim=[100], lam=0.005, lr=0.01, max_iter=50000, lam_ridge=1e-2, penalty='H', device='cuda', seed=None)[source]

Initialize cMLP

Methods

__init__([lag, hidden_dim, lam, lr, ...])

Initialize cMLP

eval(true_adj, predicted_adj[, shd_thresholds])

Evaluate the predicted adjacency matrix against the ground truth.

run(X)

Run cMLP algorithm.

run_raw(X, **kwargs)

Run the algorithm and return an unthresholded intermediate result that can be reused across multiple threshold values.

run_threshold_sweep(X, thresholds)

Run the algorithm once and post-process the raw result for each threshold.