pgmpy

Added

1. New role-aware causal graph infrastructure, including PDAG, ADMG, MAG, AncestralBase, and SimpleCausalModel.
2. New graph utilities in DAG, including to_pdag, to_dagitty, to_lavaan, public get_ancestors, edge_strength, get_stats, and __hash__.
3. DAG.to_daft and DAG.to_graphviz can now annotate plots with computed edge strengths.
4. DAG.from_dagitty can now construct LinearGaussianBayesianNetwork instances when the dagitty model includes beta coefficients.
5. New sklearn-compatible causal discovery package pgmpy.causal_discovery with refactored PC, GES, HillClimbSearch, and ExpertInLoop estimators.
6. ExpertKnowledge now supports search_space, richer string representations, temporal-order handling, and tighter integration with causal discovery and ExpertInLoop.
7. New causal effect APIs built around pgmpy.identification, including shared identification base classes and adjustment/frontdoor workflows.
8. New pgmpy.prediction module with sklearn-compatible causal prediction estimators: NaiveAdjustmentRegressor, DoubleMLRegressor, and NaiveIVRegressor.
9. New pgmpy.ci_tests package with class-based CI tests and registry-based lookup, including FisherZ, PearsonrEquivalence, and estimator-configurable GCM.
10. New pgmpy.structure_score package and new causal graph evaluation metrics: AdjacencyConfusionMatrix and OrientationConfusionMatrix.
11. LinearGaussianBayesianNetwork now supports JSON load/save, log_likelihood, predict_probability, improved predict, and richer simulation features for interventions, evidence, virtual interventions, and missing-data generation.
12. FunctionalBayesianNetwork sampling now supports vectorized FunctionalCPD functions and interventional simulation.
13. ExpectationMaximization gained apply_smoothing, configurable init_cpds, and support for node-specific priors.
14. DynamicBayesianNetwork.simulate gained a return_format argument.
15. New dataset and example-model registries, pgmpy.datasets and pgmpy.example_models, with filterable catalogs and on-demand loading.
16. Many new benchmark datasets were added, including Boston Housing, Wine Quality, Galton Stature, Seoul Bike, Pima Diabetes, Student Performance, South German Credit, Cystic Fibrosis, Yacht Hydrodynamics, HTRU2, Myocardial Infarction, IQ Brain Size, Contraceptive Method, Pittsburgh Bridges, Residential Building, Hitters, Superconductivity, Lead, Cities, College Plans, Depression Coping, US Crime, Hungary chickenpox, the Angrist-Krueger 1980 census extract, dropout covariance data, and the Tuebingen pairs.
17. Many new example models were added through bnlearn, bnrep, and DAG-only registries, including Andes, Barley, Child, Hailfinder, Hepar2, Insurance, Link, Munin 1-4, Pathfinder, Pigs, Sachs, Survey, Water, and continuous models such as magic_irri.
18. DiscreteBayesianNetwork / readwrite support now includes NET/HUGIN files, improved XDSL support, and a JSON schema for Linear Gaussian model serialization.

Changed

1. PC, GES, HillClimbSearch, and ExpertInLoop were refactored into sklearn-style causal discovery estimators with fit, learned causal_graph_, and automatic CI-test / structure-score selection based on the data type.
2. Graph role names are now standardized to plural keys such as exposures, outcomes, and confounders across graph, identification, and prediction APIs.
3. ExpertKnowledge now lives under pgmpy.causal_discovery, and structure scores are organized under pgmpy.structure_score.
4. Example datasets and example models are now loaded through Hugging Face-backed registries instead of being bundled directly in the repository.
5. MarkovNetwork is now deprecated in favor of DiscreteMarkovNetwork.
6. DAG.fit moved to DiscreteBayesianNetwork.fit, and Bayesian-network simulation argument ordering was made more consistent across model classes.
7. Causal discovery estimators can now score themselves, and structure scoring defaults are more automatic and BIC-oriented.
8. CI-test internals were refactored toward registry/class-based implementations, and scikit-base is now used for object lookup in datasets, metrics, example models, and CI tests.
9. Optional dependencies such as torch / pyro, Graphviz, documentation tooling, and LLM integrations are isolated more cleanly as extras.
10. Packaging migrated from setup.py to pyproject.toml, and pyparsing is now an explicit dependency for read/write features.
11. list_models and list_datasets now validate filter tags explicitly and expose richer metadata such as has_missing_data and has_index_col.
12. BIF, XMLBIF, NET, and XDSL writers now warn when state names contain commas, and BIFWriter no longer silently rewrites those state names.

Fixed

1. Fixes correctness issues in GES, including the latest algorithmic fix in #3228.
2. Fixes order-independence issues in the stable variant of PC, along with earlier bugs around max_cond_vars propagation and skeleton building.
3. Fixes role validation and pretreatment handling in causal prediction / DML workflows, plus multiple exposure/outcome naming inconsistencies in identification and examples.
4. Fixes random model generation edge cases in LinearGaussianBayesianNetwork, including handling of isolated latent nodes and seed=0 in get_random_cpds.
5. Fixes LinearGaussianBayesianNetwork.predict, predict_probability, invalid virtual evidence errors, unbiased standard deviation estimation, and several simulation / documentation mismatches.
6. Fixes inference and sampling issues in FunctionalBayesianNetwork, ApproxInference, and DBNInference.
7. Improves ExpectationMaximization handling of missing data by treating all-missing columns as latent variables, dropping partially missing rows explicitly, and adding clearer validation.
8. Improves BIF/XDSL/NET/XMLBIF read-write robustness, including faster and more reliable BIF parsing, better node/state-name handling, and cleaner warnings for punctuation in state names.
9. Fixes SHD on graphs with isolated nodes, CorrelationScore on very small graphs, MarkovChain.add_transition_model validation, FactorGraph.add_edge defaults, and Bayesian-estimator model reconstruction with isolated nodes.
10. Fixes compatibility issues with newer dependency and platform versions, including pandas 3 string dtype changes, newer xgboost, Python 3.14, and macOS / Windows CI environments.
11. Many notebooks, docstrings, and usage examples across causal discovery, inference, factors, and model APIs were corrected to match the current behavior.

Added

1. Option to specify the node names in random model generation methods.
2. ExpertInLoop.estimate method now accepts any LLM model supported by litellm.
3. AIC and BIC structure scoring methods for Gaussian variables and Conditional Gaussian variables.
4. Greedy Equivalence Search algorithm for causal discovery.
5. Causal Discovery algorithms can automatically figure out the data types of the variables.
6. Support for continuous and mixed data types for all causal discovery / structure learning algorithms.
7. Adds LinearGaussianBayesianNetwork and FunctionalBayesianNetwork classes to represent Gaussian and hybrid Bayesian Networks.
8. Add pgmpy.metrics.SHD to compute the Structural Hamming Distance between two DAGs.
9. Adds NoisyORCPD class to represent NoisyOr models.
10. BayesianNetwork.simulate method can not simulate different types of missing data.
11. BayesianNetwork.predict now predicts any missing value in the dataframe instead of missing columns.
12. Adds continuous example models from bnlearn repository.
13. Adds ExpertKnowledge class to specify expert knowledge for structure learning algorithms.
14. Option to initialize DAG and DiscreteBayesianNetwork with adjacency matrix, a dagitty model, or a lavaan model string.
15. Adds method for reading and writing XDSL file format (used by GeNIe).
16. Adds Generalized Covariance Measure (GCM) conditional independence test.

Fixed

1. Fixes bug in pgmpy.estimators.SEMEstimator.

Changed

1. Renames pgmpy.estimators.CITests.ci_pillai to pgmpy.estimators.CITests.pillai_trace.
2. BayesianNetwork.fit method moved to DAG.fit so that fitting can be done on either model types.
3. All structure scoring methods have been renamed to simplify.

Removed

1. Removes some of the CI test variants of chi-squared test.
2. Removes pgmpy.factors.continuous.ContinuousFactor class.
3. Removes discretization methods for ContinuousFactor.
4. BayesianModel and MarkovModel classes have been removed.
5. BayesianNetwork class have been removed. Use DiscreteBayesianNetwork instead.

Added

1. Support for returning Belief Propagation messages in Factor Graph BP.
2. Maximum Likelihood Estimator for Junction Tree.
3. Adds a simple discretization method: pgmpy.utils.discretize.
4. Two new metrics for model testing: pgmpy.metrics.implied_cis and pgmpy.metrics.fisher_c.
5. Support for Linear Gaussian Bayesian Networks: estimation, prediction, simulation and random model generation.
6. New mixed data Conditional Independence test based on canonical correlations.
7. New LLM based structure learning / causal discovery algorithm. Also LLM based pairwise variable orientation method.

Fixed

1. Reading and Writing from XBN file format.
2. Documentation for plotting models.
3. Fixes PC algorithm to add disconnected nodes in the final model.
4. Allows . in variables names in BIF file format.

Changed

1. Allows virtual_evidence parameter in inference methods to accept DiscreteFactor objects.

Added

1. init_cpds argument to ExpecattionMaximiation.get_parameters to specify initialization values.
2. BeliefPropagation with message passing for Factor Graphs.
3. Marginal Inference for undirected graphs.

Fixed

1. Incompatibality with networkx==3.2.
2. CausalInference.get_minimal_adjustment_set to accept string variable names.
3. Bug in EM when latent variables are present.
4. compat_fns.copy to consider the case when int or float is passed.
5. Fixes issue with BayesianNetwork.fit_update when running with CUDA backend.

Changed

1. Documentation Updates
2. Optimizations for Hill Climb Search algorithm.
3. Tests shutdown parallel workers in teardown.
4. Removes the complete_samples_only argument from BaseEstimator.state_counts.
5. Default number of cores to use changed to 1 for parameter estimation methods.

Added

1. Added support for python 3.11.
2. Adds DAG.to_graphviz and PDAG.to_graphviz methods to convert model to graphviz objects.
3. Adds pytorch as an alternative backend.
4. Adds unicode support for BIFReader.

Fixed

1. Warnings use a logger instance.
2. Fixes documentation.
3. Fixes variables name arguments for CausalInference.get_minimal_adjustment_set

Changed

1. Adds argument to specify samples for ApproxInference.
2. Memory optimizations for computing structure scores.
3. Switches joblib backed to loky.
4. Runtime optimizations for sampling.
5. Runtime optimizations for Variable Elimination.
6. All config variables moved to pgmpy.global_vars.

Added

1. BIFReader made compatible with the output of PyAgrum
2. Support for all available CI tests in PC algorithm.
3. References for read/write file formats.

Removed

1. Removes DAG.to_pdag method.

Changed

1. Fixes for ApproxInference for DBNs.
2. Make xml.etree the default parser instead of using lxml.

Added

1. AIC score metric from score based structure learning.
2. Adds support for NET (HUGIN) file format.
3. Adds argument reindex to state_counts method.

Fixed

1. Bug in GibbsSampling when sampling from Bayesian Networks.
2. Fix seed for all simulation methods.
3. Memory leaks when using lru_cache.

Changed

1. Caching disabled for computing state name counts during structure learning.
2. Pre-computation for sampling methods are optimized.

Added

1. BayesianNetwork.get_state_probability method to compute the probability of a given evidence.
2. BayesianEstimator.estimate_cpd accepts weighted datasets.

Fixed

1. Fixes bug in CausalInference.estimate_ate with front-door criterion.
2. Fixes inference bugs when variable has a single state.

Added

1. BayesianNetwork.get_random_cpds method to randomly parameterize a network structure.
2. Faster Variable Elimination using tensor contraction.
3. factors.factor_sum_product method for faster sum-product operations using tensor contraction.

Fixed

1. Bug in DynamicBayesianNetwork.initialize_initial_state. #1564
2. Bug in factors.factor_product. #1565

Changed

1. Runtime improvements in DiscreteFactor.marginalize and DiscreteFactor.copy methods.

Added

1. Adds checks for arguments to BayesianNetwork.simulate method.

Fixed

1. Fixes TAN algorithm to use conditional information metric.
2. Speed ups for all estimation and inference methods.
3. Fix in stable variant of PC algorithm to give reproducible results.
4. Fix in GibbsSampling for it to work with variables with integral names.
5. DAG.active_trail_nodes allows tuples as variable names.
6. Fixes CPD and edge creation in UAIReader.

Fixed

1. Fixes CausalInference.is_valid_backdoor_adjustment_set to accept str arguments for Z.
2. Fixes BayesianNetwork.remove_cpd to work with integral node names.
3. Fixes MPLP.map_query to return the variable states instead of probability values.
4. Fixes BIFWriter to generate output in standard BIF format.

Added

1. Adds BayesianNetwork.states property to store states of all the variables.
2. Adds extra checks in check model for state names

Fixed

1. Fixes typos in BayesianModel deprecation warning
2. Bug fix in printing Linear Gaussian CPD
3. Update example notebooks to work on latest dev.

Added

1. Adds a fit_update method to BayesianNetwork for updating model using new data.
2. Adds simulate method to BayesianNetwork and DynamicBayesianNetwork to simulated data under different conditions.
3. Adds DynamicBayesianNetwork.fit method to learn model parameters from data.
4. ApproxInference class to do approximate inference on models using sampling.
5. Robust tests for all sampling methods.
6. Adds BayesianNetwork.load and BayesianNetwork.save to quickly read and write files.

Changed

1. BayesianModel and MarkovModel renamed to BayesianNetwork and MarkovNetwork respectively.
2. The default value of node position in DAG.to_daft method.
3. Documentation updated on the website.

Fixed

1. Fixes bug in DAG.is_iequivalent method.
2. Automatically truncate table when CPD is too large.
3. Auto-adjustment of probability values when they don't exactly sum to 1.
4. tqdm works both in notebooks and terminal.
5. Fixes bug in CausalInference.query method.

Added

1. Adds network pruning for inference algorithms to reduce the size of network before running inference.
2. Adds support for latent variables in DAG and BayesianModel.
3. Parallel implementation for parameter estimation algorithms.
4. Adds DAG.get_random and BayesianModel.get_random methods to be able to generate random models.
5. Adds CausalInference.query method for doing do operation inference with or without adjustment sets.
6. Adds functionality to treesearch to do auto root and class node selection (#1418)
7. Adds option to specify virtual evidence in bayesian network inference.
8. Adds Expectation-Maximization (EM) algorithm for parameter estimation in latent variable models.
9. Add BDeuScore as another option for structure score when using HillClimbSearch.
10. Adds CausalInference.get_minimal_adjustment_set` for finding adjustment sets.

Changed

1. Renames DAG.is_active_trail to is_dconnected.
2. DAG.do can accept multiple variables in the argument.
3. Optimizes sampling methods.
4. CI moved from travis and appveyor to github actions.
5. Drops support for python 3.6. Requires 3.7+.

Fixed

1. Example model files were not getting included in the pypi and conda packages.
2. The order of values returned by CI tests was wrong. #1403
3. Adjusted and normalized MI wasn't working properly in TreeSearch.
4. #1423: Value error in bayesian estimation.
5. Fixes bug in DiscreteFactor.__eq__ to also consider the state names order.

Added

1. Adds support for python 3.9.
2. BayesianModelProbability class for calculating pmf for BNs.
3. BayesianModel.predict has a new argument stochastic which returns stochastic results instead of MAP.
4. Adds new method pgmpy.base.DAG.to_daft to easily convert models into publishable plots.

Changed

1. pgmpy.utils.get_example_model now doesn't need internet connection to work. Files moved locally.

Fixed

1. Latex output of pgmpy.DAG.get_independencies.
2. Bug fix in PC algorithm as it was skipping some combinations.
3. Error in sampling because of seed not correctly set.

Added

1. PC estimator with original, stable, and parallel variants.
2. PDAG class to represent partially directed DAGs.
3. pgmpy.utils.get_example_model function to fetch models from bnlearn repository.
4. Refactor HillClimbSearch with a new feature to specify fixed edges in the model.
5. Adds a global SHOW_PROGRESS variable.
6. Adds Chow-Liu structure learning algorithm.
7. Add pgmpy.utils.get_example_model to fetch models from bnlearn's repository.
8. Adds get_value and set_value method to DiscreteFactor to get/set a single value.
9. Adds get_acestral_graph to DAG.

Changed

1. Refactors ConstraintBasedEstimators into PC with a lot of general improvements.
2. Improved (faster, new arguments) independence tests with changes in argument.
3. Refactors sample_discrete method. Sampling algorithms much faster.
4. Refactors HillClimbSearch to be faster.
5. Sampling methods now return dataframe of type categorical.

Fixed

Removed

1. Data class.

Added

• New example notebook: Alarm.ipynb
• Support for python 3.8
• Score Caching support for scoring methods.

Changed

• Code quality check moved to codacy from landscape
• Additional parameter max_ci_vars for ConstraintBasedEstimator.
• Additional parameter pseudo_count for K2 score.
• Sampling methods return state names instead of number when available.
• XMLBIFReader and BIFReader not accepts argument for specifying state name type.

Fixed

• Additional checks for TabularCPD values shape.
• DiscreteFactor.reduce accepts both state names and state numbers for variables.
• BeliefPropagation.query fixed to return normalized CPDs.
• Bug in flip operation in HillClimbSearch.
• BIFWriter to write the state names to file if available.
• BayesianModel.to_markov_model fixed to work with disconnected graphs.
• VariableElimination fixed to not ignore identifical factors.
• Fixes automatic sorting of state names in estimators.

Removed

• No support for ProbModelXML file format.