Metrics#

After learning a graph or fitting a model, you need to evaluate how good the result is. pgmpy provides metrics for two settings: supervised evaluation compares a learned graph against a known ground-truth graph, and unsupervised evaluation scores a graph against the observed data when no ground truth is available.

At a Glance#

Unified API: All metrics share a consistent callable interface.
Supervised Metrics: Compare learned graphs against ground truth using structural distance and confusion matrices.
Unsupervised Metrics: Score graphs against data using correlation residuals, implied conditional independencies, and structure scores.
Metric Discovery: Programmatically find available metrics by their requirements.

API#

All metrics follow the same pattern — instantiate and call:

from pgmpy.base import DAG
from pgmpy.metrics import SHD

true_graph = DAG([("A", "B"), ("B", "C")])
estimated_graph = DAG([("B", "A"), ("B", "C")])

metric = SHD()
print(metric(true_causal_graph=true_graph, est_causal_graph=estimated_graph))

Supervised Metrics#

When a ground-truth graph is available, supervised metrics measure the structural distance between the estimated and true graphs — counting edge additions, deletions, and reversals, or computing confusion matrices for adjacency and orientation accuracy.

Unsupervised Metrics#

When no ground truth is available, unsupervised metrics score how well a graph explains the observed data. These check whether the conditional independencies implied by the graph hold in the data, or evaluate the graph under a configurable structure scoring method.

Metric Discovery#

get_metrics(...) lets you discover available metrics by their requirements instead of hard-coding class names:

from pgmpy.metrics import get_metrics

supervised = get_metrics(requires_true_graph=True)
unsupervised = get_metrics(requires_data=True)

API Reference#

For the full list of supported metrics: