Source code for dodiscover.toporder.score

from typing import List, Optional, Tuple

import numpy as np
from numpy.typing import NDArray

from dodiscover.toporder._base import BaseTopOrder, SteinMixin
from dodiscover.toporder.utils import full_dag, pns


[docs]class SCORE(BaseTopOrder, SteinMixin): """The SCORE algorithm for causal discovery. SCORE :footcite:`rolland2022` iteratively defines a topological ordering finding leaf nodes by comparison of the variance terms of the diagonal entries of the Hessian of the log likelihood matrix. Then it prunes the fully connected DAG representation of the ordering with CAM pruning :footcite:`Buhlmann2013`. The method assumes Additive Noise Model and Gaussianity of the noise terms. Parameters ---------- eta_G: float, optional Regularization parameter for Stein gradient estimator, default is 0.001. eta_H : float, optional Regularization parameter for Stein Hessian estimator, default is 0.001. alpha : float, optional Alpha cutoff value for variable selection with hypothesis testing over regression coefficients, default is 0.05. prune : bool, optional If True (default), apply CAM-pruning after finding the topological order. n_splines : int, optional Number of splines to use for the feature function, default is 10. Automatically decreased in case of insufficient samples splines_degree: int, optional Order of spline to use for the feature function, default is 3. estimate_variance : bool, optional If True, store estimates the variance of the noise terms from the diagonal of Stein Hessian estimator. Default is False. pns : bool, optional If True, perform Preliminary Neighbour Search (PNS) before CAM pruning step, default is False. Allows scaling CAM pruning and ordering to large graphs. pns_num_neighbors: int, optional Number of neighbors to use for PNS. If None (default) use all variables. pns_threshold: float, optional Threshold to use for PNS, default is 1. References ---------- .. footbibliography:: Notes ----- Prior knowledge about the included and excluded directed edges in the output DAG is supported. It is not possible to provide explicit constraints on the relative positions of nodes in the topological ordering. However, explicitly including a directed edge in the DAG defines an implicit constraint on the relative position of the nodes in the topological ordering (i.e. if directed edge `(i,j)` is encoded in the graph, node `i` will precede node `j` in the output order). """ def __init__( self, eta_G: float = 0.001, eta_H: float = 0.001, alpha: float = 0.05, prune: bool = True, n_splines: int = 10, splines_degree: int = 3, estimate_variance=False, pns: bool = False, pns_num_neighbors: Optional[int] = None, pns_threshold: float = 1, ): super().__init__( alpha, prune, n_splines, splines_degree, pns, pns_num_neighbors, pns_threshold ) self.eta_G = eta_G self.eta_H = eta_H self.var: List[float] = list() # data structure for estimated variance of SCM noise terms self.estimate_variance = estimate_variance def _top_order(self, X: NDArray) -> Tuple[NDArray, List[int]]: """Find the topological ordering of the causal variables from X dataset. Parameters ---------- X : np.ndarray Dataset of observations of the causal variables Returns ------- A_dense : np.ndarray Fully connected matrix admitted by the topological ordering order : List[int] Inferred causal order """ def _estimate_var(stein_instance, data): H_diag = stein_instance.hessian_diagonal(data, self.eta_G, self.eta_H) H_diag = H_diag / H_diag.mean(axis=0) return 1 / H_diag[:, 0].var(axis=0).item() _, d = X.shape order: List[int] = list() active_nodes = list(range(d)) stein = SteinMixin() for _ in range(d - 1): X = self._find_leaf_iteration(stein, X, active_nodes, order) order.append(active_nodes[0]) if self.estimate_variance: self.var.append(_estimate_var(stein, X)) order.reverse() return full_dag(order), order def _prune(self, X: NDArray, A_dense: NDArray) -> NDArray: """Pruning of the fully connected adj. matrix representation of the inferred order. If self.do_pns = True or self.do_pns is None and number of nodes >= 20, then Preliminary Neighbors Search is applied before CAM pruning. Parameters ---------- X : np.ndarray of shape (n_samples, n_nodes) Matrix of the data. A_dense : np.ndarray of shape (n_nodes, n_nodes) Dense adjacency matrix to be pruned. Returns ------- A : np.ndarray The pruned adjacency matrix output of the causal discovery algorithm. """ d = A_dense.shape[0] if (self.do_pns) or (self.do_pns is None and d > 20): A_dense = pns( A=A_dense, X=X, pns_threshold=self.pns_threshold, pns_num_neighbors=self.pns_num_neighbors, ) return super()._prune(X, A_dense) def _find_leaf_iteration( self, stein_instance: SteinMixin, X: NDArray, active_nodes: List[int], order: List[int] ) -> NDArray: """Find a leaf by inspection of the diagonal elements of the Hessian of the log-likelihood. Leaf node equals the 'argmin' of the variance of the diagonal terms of the Hessian of the log-likelihood. After a leaf is identified, it is added to the topological order, and the list of nodes without position in the ordering is updated. Parameters ---------- stein_instance : SteinMixin Instance of the Stein estimator with helper methods. X : np.ndarray Matrix of the data (active nodes only). active_nodes : List[int] List of the nodes without a position in the topological ordering. order : List[int] List of partial order. Returns ------- X : np.ndarray Matrix of the data without the column corresponding to the identified leaf node. """ H_diag = stein_instance.hessian_diagonal(X, self.eta_G, self.eta_H) leaf = int(H_diag.var(axis=0).argmin()) leaf = self._get_leaf( leaf, active_nodes, order ) # get leaf consistent with order imposed by included_edges from self.context order.append(active_nodes[leaf]) active_nodes.pop(leaf) X = np.hstack([X[:, 0:leaf], X[:, leaf + 1 :]]) if self.estimate_variance: self.var.append(1 / H_diag[:, leaf].var(axis=0).item()) return X