This major release brings three headline changes: the causal module is promoted from pure Python to a first-class C++
module (CM), structure learning is extended with PC, FCI (with PAG output), GreedyThickThinning, and triangle
deletions in GreedyHillClimbing, and a new native GUM serialization format (jgum / bgum) is introduced for all
graphical models. It also consolidates widespread API modernisation (C++20 std::format, string_view, optional_ref,
std::optional, concepts) and several breaking renames detailed below.

• pyAgrum

  • Breaking API:
    • Graph path-finding functions undirectedPath, directedPath, directedUnorientedPath,
      mixedOrientedPath, mixedUnorientedPath now return None instead of raising NotFound (or returning [])
      when no path exists.
    • Causal module renames: CausalModel.observedBN() → observationalBN();
      CausalModel.addLatentVariable(..., keepArcs=...) parameter keepArcs renamed to
      assumeNonSpurious; CausalModel.backDoor()/frontDoor() return None instead of raising when no valid set
      exists (set() when the empty set is valid); DoorCriteria's EnumerationOptions
      removed (options excluded_nodes, max_cardinality, only_minimal, stopAtFirst are now direct keyword
      parameters, all methods now static); CausalImpact direct access to result removed (use
      impact()); Counterfactual.getResult() renamed impact(); DSeparation renamed Separation.
    • NodeId and NameOrId type aliases removed from pyagrum.ctbn and pyagrum.clg (use int and
      int | str).
    • _gum_add_properties_while_getstate_ removed; pickle metadata now handled by updateMetaData()
      (pickled objects from earlier versions may not unpickle correctly).
    • BNClassifier now requires a DiscreteTypeProcessor constructor parameter; createBNClassifier
      signature changed accordingly; binary prediction dispatch and predict_proba updated; model()
      method added (returns a copy of the underlying BN).
  • Causal Module:
    • Added SWIG Python bindings for the C++ causal module (CausalModel, CausalFormula,
      DoorCriteria, Counterfactual, ...).
    • Refactored causalEffectEstimation as a top-level subpackage.
    • Fixed type references in docstrings (pyagrum.causal.CausalModel → pyagrum.CausalModel) and restructured
      the causality section in the Sphinx documentation.
    • Updated causality notebooks; added causal2graph utility for causal graph visualization; added AST printing
      for CausalImpact results.
  • Structure Learning (FCI / PAG):
    • Added SWIG bindings for PAG (Partial Ancestral Graph) and the FCI (Fast Causal Inference)
      algorithm.
    • Exposed BNLearner.setAlgorithmFCI(), fciExhaustiveSepSet() / setFCIExhaustiveSepSet().
    • Added documentation and test suite for PAG/FCI.
  • Native GUM Format (jgum / bgum):
    • Added saveGUM(), loadGUM(), saveGUMstring(), loadGUMstring() to BayesNet,
      InfluenceDiagram, and MarkovRandomField.
    • Added SWIG type mapping for optional<vector<NodeId>>.
    • Added GumFormatTestSuite covering jgum/bgum round-trips and string serialization.
    • Added Sphinx page jgum-bgum-format.rst and notebook 91-Tools-LoadAndSaveGraphicalModels
      illustrating all I/O formats.
  • qBNSampling (experimental, thanks to Tibor Dubois, Thierry Rioual, Mehmet Gunes):
    • New pyagrum.qBNSampling module: quantum circuit encoding of Bayesian Networks and rejection-sampling
      inference.
  • CLG (Conditional Linear Gaussian):
    • Added CLG.asDiscreteBN() to convert a CLG model to a discretized BN.
    • Added CLG.__eq__, __getstate__, __setstate__ (pickle support).
    • New CLGModelTestSuite and extended CLGInference/SEM/Sampling/Random test suites.
    • randomCLG gains max_parents and ratio_arc parameters; GaussianVariable sigma guard added (thanks to
      Ima Bernada).
  • BNClassifier & skbn (thanks to Lou Toubiana):
    • Added model() method returning a copy of the underlying BN; full coverage
      BNClassifierTestSuite.
    • Refactored skbn for full sklearn API compliance: ClassifierMixin placed before
      BaseEstimator in MRO, fitted attributes renamed to the trailing-underscore convention (bn_,
      threshold_, target_, ...), validate_data() used for input validation in fit, predict, and
      predict_proba.
    • fit(X, y) signature is now positional and strictly sklearn-compliant; the old
      fit(X=None, y=None, data=None, targetName=None) form is removed.
    • Fixed pandas 2.x compatibility (X cast to object dtype before processing in fit) and suppressed a
      spurious sklearn 'X does not have valid feature names' warning in predict /
      predict_proba.
  • StructuralMetrics:
    • Exposed StructuralMetrics (formerly StructuralComparator) with SHD, tp/fp/fn/tn accessors and SID
      (Structural Intervention Distance) for BN comparison.
    • Updated GraphicalBNComparator to delegate to StructuralMetrics; added structuralFScore.
  • Tensor & Numpy:
    • Tensor methods random(), randomDistribution(), randomCPT(), noising() now return self
      for chaining.
    • Added Python bindings for mean(), variance(), stdDev(), isNumerical().
    • Numpy interop: new as_nparray(), toarray(), and fillWith(ndarray) methods; optimized
      Tensor numpy access with zero-copy __getitem__ and memcpy-based __setitem__ (numpy >= 1.7 required).
  • Graph API:
    • Exposed nameFromId(), idFromName(), setName(), hasName() on all graph types (DiGraph,
      DAG, MixedGraph, ...).
    • Added connectedComponents() (returns dict[int, int]), connectedComponentsList(), and
      connectedComponentsCount().
    • Fixed missing descendants() / ancestors() on DiGraph; NodeSet, ArcSet, EdgeSet now use dedicated
      typemaps.
  • Documentation:
    • Added sentinel typedefs to replace -> object with precise Python return type annotations in generated
      bindings.
    • Added %feature(docstring) entries for all 2174 methods; docstrings follow NumPy format (100% docstring
      coverage).
  • Performance & Infrastructure:
    • Introduced PYTHONIZED_MARGINALS macro and global numpy import for generated inference code.
    • Lazy-import of pandas via TYPE_CHECKING (~240 ms saved on import time).
    • Added -fvisibility=hidden: _pyagrum.so reduced from 16.8 MB to 12.9 MB (−23 %).
    • Added full type annotations to all pyLibs (lib/, clg/, skbn/, bnmixture/, ctbn/,
      causalEffectEstimation/); added inference type aliases BNInference, MRFInference,
      CNInference; replaced mypy with pyrefly in act guideline, fixing all type errors across pyLibs;
      added MatrixLike/ArrayLike type aliases to __init__.in.py.
    • BNMixture: manifest-based saveBNM/loadBNM, loadRetroCompatibleBNM for backward compatibility; fixed
      saveBNM/loadBNM on Windows (colons in BN names).
    • explain (ShAP / SHALL): API improvements, causal SHAP fix, binary BN test resources added; fixed
      uninitialized array in _labelToPos (np.empty → np.zeros).
    • Rewritten ipython.py, new jt2graph.py, fixed prepareDot fontcolor; fixed html2image
      issues in notebook contexts and improved export cropping.
    • act install: support no-make mode via cmake --install; --only flag (alias for --build no-make). act test pyAgrum: persistent --test_build_path option. Fixed fastPrototype
      docstring separator; removed useless newFactory methods.

• aGrUM

  • Breaking API:
    • Graph path-finding functions directedPath, directedUnorientedPath, undirectedPath,
      mixedOrientedPath, mixedUnorientedPath now return
      std::optional<std::vector<NodeId>> (nullopt when no path exists), for DiGraph, UndiGraph,
      MixedGraph, PDAG, CliqueGraph, and all GUM_DiGraphable/GUM_UndiGraphable/
      GUM_MixedGraphable graphs.
    • Causal module renames: CausalModel::observedBN() → observationalBN();
      CausalModel::addLatentVariable(..., keepArcs=...) parameter keepArcs renamed to
      assumeNonSpurious; CausalModel::backDoor()/frontDoor() return type NodeSet →
      std::optional<NodeSet> (nullopt = no valid set, {} = empty set is valid); DoorCriteria's
      EnumerationOptions struct removed (options are now direct parameters, all methods now
      static), DoorCriteria::nodesOnDirectedPaths(dag, X, Y) return type NodeSet →
      std::optional<NodeSet>; CausalImpact direct access to result removed (use impact());
      Counterfactual<GUM_ELEMENT> template parameter renamed from GUM_SCALAR, getResult() renamed
      impact(); DSeparation (in CM/tools/) renamed Separation.
    • HashTable<Key,Val>::tryGet(key) and related methods (tryFirst, trySecond, tryPos) now return
      optional_ref<Val> / optional_ref<const Val> instead of Val* / const Val*
      (nullptr if absent). gum::optional_ref<T> behaves like std::optional for references (C++26 feature
      backported); callers using if (auto* p = table.tryGet(key)) must switch to
      if (auto ref = table.tryGet(key)).
    • Signaler1<A>, Signaler2<A,B>, Signaler3<A,B,C>, ... removed; use variadic Signaler<A>,
      Signaler<A,B>, Signaler<A,B,C>, ... instead.
    • Widespread string_view migration: const std::string& parameters replaced by
      std::string_view across the public API (BayesNet, variables, I/O readers, learning, ...); passing
      std::string or string literals remains compatible.
    • I/O writers' write() no longer const: BNWriter<GUM_SCALAR>::write() (and all subclass writers, including
      GumBNWriter) now takes the BN by non-const reference, to allow writers to call bn.updateMetaData()
      before serialization; CredalNet::saveBNsMinMax() is similarly affected.
    • New base class DiscreteGraphicalModel: IBayesNet and IMarkovRandomField now inherit from it, which
      factors five variable-map accessors (variable, variableNodeMap, nodeId,
      idFromName, existsInModel) previously duplicated in each interface; direct subclasses must no longer
      define those accessors themselves.
    • DiscreteVariable::closestLabel is now virtual (ABI change: recompilation required for any code linking
      against aGrUM as a shared library).
    • DAGmodel::dag() and UGmodel::graph() now return value copies (with node names propagated), not const
      references; use internalDag() / internalGraph() for the O (1) stable const
      reference when graph mutation through the model API is not needed.
    • StructuralComparator renamed StructuralMetrics (update all include paths and type names accordingly).
  • Structure Learning:
    • Added the FCI (Fast Causal Inference) algorithm producing a PAG (Partial Ancestral Graph)
      from data; integrated into IBNLearner/BNLearner (setAlgorithmFCI); fixed possibleDSep
      criterion (Zhang 2008); added exhaustive sepset mode (setFCIExhaustiveSepSet); enforced background knowledge
      in orientation rules (R1/R2/R9/R10).
    • Added the PC constraint-based structure learning algorithm, integrated into
      IBNLearner/BNLearner alongside Miic.
    • Added the GreedyThickThinning score-based structure learning algorithm, integrated into
      IBNLearner/BNLearner.
    • Extended GreedyHillClimbing with arc-triangle deletion operations:
      GraphChangesSelector4DiGraph now supports applyArcDeletion, applyArcReversal,
      applyTriangleDeletion; LocalSearchWithTabuList updated; totalOrder constraint added.
    • Independence tests refactoring: extracted CachedContingencyCounter base class from Chi2 and G2
      implementations; IndependenceTest::statistics() made pure virtual (overridden in
      IndepTestChi2 and IndepTestG2); added silent-cell df correction in Chi2/G2 tests; fixed G2 df for sampling
      zeros; new Chi2TestSuite.
    • ConstraintBasedLearning refactoring: extracted ConstraintBasedLearning base class from Miic; extracted
      CIBasedLearning base class adding scorer-agnostic API (learnPDAG/learnDAG/learnBN); CMI types and
      comparators moved to Miic;
      setMutualInformation() injector added; applyStructuralConstraints_ factored into
      ConstraintBasedLearning; renamed scores_and_tests/ directory to scores/.
  • Causal Module (CM) Development (thanks to SCALNYX):
    • Promoted the causal module from pure Python to a first-class C++ module.
    • Introduced CausalModel and CausalFormula.
    • Developed an Abstract Syntax Tree (AST) for do-calculus, including LaTeX export and evaluation.
    • Added the DoorCriteria class for backdoor and frontdoor set enumeration.
    • Implemented ID/IDC algorithms and formula introspection.
    • Added counterfactual and counterfactualModel functions with associated tests.
  • Native GUM Format (jgum / bgum):
    • Added GumBNReader/GumBNWriter, GumIDReader/GumIDWriter, GumMRFReader/GumMRFWriter
      supporting both JSON (.jgum) and binary (.bgum) serialization.
    • All GUM readers support proceedFromString() and a no-filename constructor for in-memory round-trips.
    • Extracted _readVector_/_writeVector_ helpers to GumBinaryIO.h.
    • Fixed empty-BN jgum serialization; fixed binary writers to open files with ios::binary; fixed
      IDReader::proceed() return type (void → Size).
  • Modeling & Core API:
    • Added optional node name support to NodeGraphPart: nameFromId(), idFromName(),
      setName(), hasName(); names propagated to toDot() output (format: id:name);
      checkConsistency() made public, friend declarations for test suites removed;
      GraphicalModel::_nameNodes_() propagates node names to returned graphs (moralGraph(),
      moralizedAncestralGraph(), EssentialGraph::pdag() / skeleton(), MarkovBlanket::dag()).
    • Introduced DiscreteGraphicalModel to factorize variable management across IBayesNet and
      IMarkovRandomField.
    • Added connectedComponents() to DAGmodel and UGmodel, with C++ and Python tests for BN, ID, MRF; added
      descendants() and ancestors() to MarkovBlanket and EssentialGraph.
    • Replaced nullable pointers with optional_ref<T> and implemented std::optional in various interfaces; added
      a data() method to MultiDimArray for contiguous buffer access.
    • Made DiscreteVariable::closestLabel(double) virtual; added DiscreteVariable::isNumerical()
      (returns true iff varType != LABELIZED).
    • Added Tensor::mean(), variance(), stdDev() (fixes variance computation: was E[X²], now
      E[(X−μ)²]); added isCloseToZero() / isCloseToOne() helpers used in these methods;
      Tensor::toString now uses Unicode box-drawing characters (│ ║ ─) for table borders.
    • StructuralMetrics (formerly StructuralComparator): added SHD metrics with tp/fp/fn/tn accessors; added SID
      (Structural Intervention Distance) for DAG-vs-DAG and BayesNet overload; name-based alignment for BN
      compare/SID.
    • BIFXMLBNReader improvements (thanks to Omi Johnson): added std::istream constructor for in-memory parsing;
      now reads the network NAME property from BIF/XML files.
    • C++20 graph concepts: added GUM_DiGraphable, GUM_UndiGraphable, GUM_MixedGraphable concepts for
      graph-agnostic programming; added generic path/reachability/cycle algorithms operating on any
      concept-satisfying graph type; added generic moralization and separation algorithms (DAG/PDAG
      methods now delegate); added generic BayesBall algorithm (dSeparated and
      BayesBall::requisiteNodes delegate); promoted minimalCondSet, markovBlanket, areConnected
      to the generic graph layer.
    • CN module: fixed critical bugs (invalid delete[] on strtok pointer, setCPT const-ref signature,
      insertEvidenceFile override placement, operator precedence in LrsWrapper guards); fixed naming convention
      violations (protected attributes renamed to trailing-underscore convention).
  • Code Quality & Static Analysis:
    • Replaced std::stringstream with std::format across all modules; added override specifier to all virtual
      method overrides; added [[nodiscard]] to clone() and factory methods; normalized include guards to the
      GUM_SOMETHING_H convention; added parent #include in _inl.h/_tpl.h
      files for IDE LSP support.
    • Integrated clang-tidy into act guideline (check tidy / --correction applies fixes); a DeepSeek
      static-analysis audit fixed all CRIT/HIGH/MED/LOW issues across the codebase;
      clang-format applied to C++ test suites and BN learning sources; fixed noexcept on
      IndepTestChi2/G2 moves and braced-init returns.
    • Portability: fixed int2Pow to use uint64_t for portable 64-bit shift on Windows; fixed GCC 16 warnings
      (-Warray-bounds pragmas, gum::Size casts for signed/unsigned comparisons); fixed GCC
      optimize pragma guard against Clang in BNLearner; fixed -Wextra-semi and extern template
      SWIG warnings; fixed Windows compilation issues (binary writers, CI uninstall).
    • Learning: MeekRules improvements; fixed propagateToCPDAG (restored edges-before-arcs insertion order in
      PDAG); fixed ArcDeletion bug in GraphChangesSelector4DiGraph; BayesBall
      (_bayesBall_) replaced the exists+insert+[] pattern with getWithDefault, plus 9 deterministic tests.
    • Migrated the test framework from CxxTest to doctest (updated from 2.4.12 to 2.5.2, suppressed
      -Wc2y-extensions); replaced GUM_CHECK_* macros with CHECK_* across all test suites; replaced deprecated
      tmpnam with a getTempFilePath helper across all test suites; added LpInterface/LrsWrapper tests and
      marginal sanity checks; removed improper try/catch
      logic, replacing it with explicit existence checks; optimized Coco/R parser performance; fixed various MSVC
      compilation issues and name lookup errors (notably regarding gum::Arc).
    • Added move constructors and move assignment operators across the class hierarchy; fixed GUM debug macros and
      the atexit table.
    • CMake: removed uninstall target and obsolete policies; AVLTree move operator= no longer
      noexcept (contains GUM_ERROR).
  • Build & Tooling:
    • act: --stats flag for project stats; --consolidate split; cm.h and base/io added to the dependency
      map; act guideline gains --dry-run, --check with +/- syntax, non-persistent
      --verbose.

Older changelogs

• Changelogs for 2.*.*
• Changelogs for 1.*.*
• Changelogs for 0.*.*