---

name: data-io-and-evaluation description: "Guide pgmpy datasets, example models, model I/O formats, and evaluation metrics." disable-model-invocation: true metadata: disco-role: operating license: MIT

Data I/O and Evaluation

Use this sub-skill when a task asks to discover or load pgmpy built-in datasets or example models, validate local data artifacts, read or write model files, or evaluate learned structures/models with pgmpy metrics.

Do not use this sub-skill to choose or run learning algorithms, construct new models/CPDs, or perform inference on a loaded model. Route structure/parameter learning to learning-structure-and-parameters, graph/CPD construction to modeling-and-factors, and posterior queries/simulation on loaded models to inference-sampling-and-simulation.

First decisions

1. Decide whether the task can stay local:
   • list_datasets(...) and list_models(...) inspect registries and should not fetch data.
   • load_dataset(...) and load_model(...) may read the local Hugging Face cache or download public assets when the cache is missing. Keep network use explicit and optional.
2. Pick the I/O layer:
   • Use model-level model.save(...) and DiscreteBayesianNetwork.load(...) for common discrete-BN files: BIF, UAI, XMLBIF, XDSL, and NET.
   • Use pgmpy.readwrite reader/writer classes for format-specific options, strings, properties, XBN, PomdpX, or round-value control.
   • Use LinearGaussianBayesianNetwork.save/load for its JSON schema.
3. Pick the metric by inputs:
   • Ground-truth graph available: use SHD, AdjacencyConfusionMatrix, or OrientationConfusionMatrix after aligning node sets.
   • No ground truth but data available: use CorrelationScore, ImpliedCIs, FisherC, or StructureScore with a data frame whose columns cover graph nodes.
4. Confirm the evaluation artifact exists. Metrics evaluate an existing graph or fitted/discovered model; they do not fit parameters or discover structures.
5. For automation, write outputs to user or temporary paths, set seeds where an API supports them, and disable progress bars on long metric checks.

Routing table

Need	Primary pgmpy entry point	Notes
List dataset names by tags	pgmpy.datasets.list_datasets(**filters)	Local registry lookup. Common filters include is_discrete, is_continuous, has_ground_truth, n_samples, and n_variables.
Load a built-in dataset	pgmpy.datasets.load_dataset(name, n_samples=None, seed=None, **sim_kwargs)	Returns a Dataset with data, ground_truth, expert_knowledge, and tags; may require cache/network except for purely simulated assets.
List example model names by tags	pgmpy.example_models.list_models(**filters)	Local registry lookup. Filters include is_parameterized, is_discrete, is_continuous, n_nodes, and n_edges.
Load an example model	pgmpy.example_models.load_model(name)	Returns a DAG, DiscreteBayesianNetwork, LinearGaussianBayesianNetwork, or optional functional model family; may require cache/network.
Persist a discrete BN	model.save(path, filetype=...) and DiscreteBayesianNetwork.load(path, filetype=...)	Supports BIF, UAI, XMLBIF, XDSL, and NET through model-level helpers.
Format-specific parsing/writing	pgmpy.readwrite.*Reader / *Writer	Use for BIF, XMLBIF, NET, UAI, XDSL, XBN, and PomdpX options or string-based reads.
Persist a linear Gaussian BN	LinearGaussianBayesianNetwork.save(path) and .load(path)	JSON file/object with nodes, arcs, and cpds; see the bundled reference for schema notes.
Supervised graph comparison	SHD, AdjacencyConfusionMatrix, OrientationConfusionMatrix	Graphs must have identical node sets. Orientation confusion supports DAGs only.
Unsupervised graph/data evaluation	CorrelationScore, ImpliedCIs, FisherC, StructureScore	Requires a pandas data frame with columns for every graph node.

Bundled references and script

• I/O, data, and metrics API map lists registry APIs, supported file formats, model-level versus class-level I/O, data artifact checks, and metric selection recipes.
• Troubleshooting covers cache/network misses, unsupported file types, malformed BIF/XML/UAI content, metric node alignment, and using metrics before fitting or discovery.
• data_io_smoke.py lists local registries, builds a tiny discrete BN, writes/reads a temporary BIF file, and computes graph metrics without reading repository fixtures or using network by default.

Minimal copyable pattern

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

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

# Supervised metrics require identical node sets, including isolated nodes.
est_graph.add_nodes_from(true_graph.nodes())
print(SHD()(true_causal_graph=true_graph, est_causal_graph=est_graph))

If that fails, inspect the exact graph node sets before changing the metric. Most metric errors come from missing/extra nodes, using a PDAG with a DAG-only metric, or passing data whose columns do not cover the graph variables.