Model learning
Learn probabilistic model structure and parameters from data.
Start with the modeling task: discover a candidate structure, fit parameters to a declared structure, or run a family-specific combined procedure. Choose from the complete catalog of supported learning algorithms below.
Choose the learning job
Structure, parameters, or a specialized combination
Structure learning
Discover a bounded candidate dependency structure from supported observations. A route is one supported variant of an algorithm for a particular model family and version. “2 routes” or “4 routes” means two or four supported variants—not steps you need to complete.
22 algorithms
- Naive Bayes (opens background reference in a new tab)Build a classifier with the class variable as the parent of each feature. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 2 routes
- Maximum-weight spanning tree (MWST) (opens background reference in a new tab)Recover a tree-structured dependency skeleton from pairwise association scores. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 2 routes
- Peter–Clark (PC) algorithm (opens background reference in a new tab)Use conditional-independence tests to recover a causal equivalence-class structure. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 4 routes
- Three-phase dependency analysis (TPDA) (opens background reference in a new tab)Discover and orient dependencies through a staged constraint-based search. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 2 routes
- Constrained three-phase dependency analysis (TPDA) (opens background reference in a new tab)Run TPDA while honoring required or forbidden structural constraints. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 2 routes
- Tree-augmented Naive Bayes (TAN) (opens background reference in a new tab)Learn a classifier whose features form a dependency tree in addition to the class parent. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 2 routes
- Bayesian network-augmented Naive Bayes (BAN) (opens background reference in a new tab)Learn a classifier with a broader feature dependency network around the class variable. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 2 routes
- Genetic algorithm (GA) search (opens background reference in a new tab)Search candidate graph structures with an evolutionary optimization strategy. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 2 routes
- K2 ordered search (opens background reference in a new tab)Build a directed acyclic graph by adding parents according to a supplied variable ordering. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 2 routes
- Hill-climbing search (opens background reference in a new tab)Improve a candidate graph through locally scored edge additions, removals, and reversals. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 2 routes
- Tabu search (opens background reference in a new tab)Explore scored graph edits while using short-term memory to escape local optima. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 3 routes
- Order-constrained search (opens background reference in a new tab)Search for a directed acyclic graph while preserving a declared variable order. Discrete Bayesian networks and influence diagrams · Linear Gaussian models · 2 routes
- Linear DAGMA (opens background reference in a new tab)Learn a linear directed acyclic graph with a continuous acyclicity optimization. Linear Gaussian models · 1 route
- DirectLiNGAM (opens background reference in a new tab)Estimate a linear non-Gaussian acyclic causal structure from observational data. Linear Gaussian models · 1 route
- Bayesian information criterion (BIC) hill climbing (opens background reference in a new tab)Use BIC-scored local search for temporal and conditional linear Gaussian structures. Conditional linear Gaussian models · Conditional linear Gaussian temporal models · Discrete temporal models · Gaussian temporal models · 4 routes
- Pairwise Bayesian information criterion (BIC) search (opens background reference in a new tab)Select pairwise factor-graph structure with penalized likelihood scores. Factor graphs · 1 route
- Exhaustive conditional-independence (CI) search (opens background reference in a new tab)Recover a finite discrete LWF chain graph through bounded conditional-independence tests. Lauritzen–Wermuth–Frydenberg (LWF) chain graphs · 1 route
- LearnSPN mixture learning (opens background reference in a new tab)Learn a probabilistic circuit by recursively partitioning variables and observations. Probabilistic circuits · 1 route
- Really Fast Causal Inference (RFCI) (opens background reference in a new tab)Recover a partial ancestral graph (PAG) while allowing latent confounding. Latent causal models · 1 route
- Ancestral directed mixed graph (ADMG) Gaussian fitting (opens background reference in a new tab)Fit the directed and bidirected structure of an ancestral linear Gaussian causal model. Latent causal models · 1 route
- Bounded Bayesian information criterion (BIC) search (opens background reference in a new tab)Learn parent sets for a continuous-time Bayesian network within declared search bounds. Continuous-time Bayesian networks (CTBNs) · 1 route
- Partial-observation expectation-maximization (EM) (opens background reference in a new tab)Learn a continuous-time Bayesian network from partially observed event histories. Continuous-time Bayesian networks (CTBNs) · 1 route
Parameter learning
Fit family-native parameters while preserving the declared model structure. A route is one supported variant of an algorithm for a particular model family and version. “2 routes” or “4 routes” means two or four supported variants—not steps you need to complete.
5 algorithms
- Supplied-structure parameter fitting (opens background reference in a new tab)Estimate family-native parameters while preserving a supplied graph or model structure. Conditional linear Gaussian models · Conditional linear Gaussian temporal models · Continuous-time Bayesian networks (CTBNs) · Discrete Bayesian networks and influence diagrams · Discrete temporal models · Factor graphs · Gaussian temporal models · Lauritzen–Wermuth–Frydenberg (LWF) chain graphs · Linear Gaussian models · 10 routes
- Supplied-structure expectation-maximization (EM) (opens background reference in a new tab)Fit probabilistic-circuit parameters with EM while preserving the supplied circuit structure. Probabilistic circuits · 1 route
- Linear Gaussian parameter fitting (opens background reference in a new tab)Estimate coefficients and noise terms for a supplied latent causal structure. Latent causal models · 1 route
- Chance-node parameter fitting (opens background reference in a new tab)Estimate chance-node probability tables for a supplied decision-analysis network. Decision-analysis networks · 1 route
- Influence-diagram chance-node fitting (opens background reference in a new tab)Estimate chance-node probability tables while preserving the supplied influence-diagram structure. Discrete Bayesian networks and influence diagrams · 1 route
Combined and specialized learning
Use family-specific procedures that jointly or iteratively fit hidden structure and parameters. A route is one supported variant of an algorithm for a particular model family and version. “2 routes” or “4 routes” means two or four supported variants—not steps you need to complete.
8 algorithms
- Hidden Markov model (HMM) Baum–Welch estimation (opens background reference in a new tab)Estimate transition and emission parameters for a hidden Markov model. Discrete temporal models · 1 route
- Linear Gaussian state-space model (LGSSM) expectation-maximization (EM) (opens background reference in a new tab)Estimate dynamics and observation parameters for a linear Gaussian state-space model. Gaussian temporal models · 1 route
- Exact maximum marginal likelihood (MML) (opens background reference in a new tab)Optimize Gaussian-process hyperparameters by exact marginal likelihood. Gaussian processes · 1 route
- Fully observed Markov decision process (MDP) fitting (opens background reference in a new tab)Estimate transition and reward mechanisms from fully observed trajectories. Partially observable Markov decision processes (POMDPs) · 1 route
- Fixed-structure POMDP expectation-maximization (EM) (opens background reference in a new tab)Estimate parameters for a partially observable Markov decision process with declared structure. Partially observable Markov decision processes (POMDPs) · 1 route
- Global intervention calculus when the DAG is absent (IDA) (opens background reference in a new tab)Estimate a set of possible total effects across graphs represented by an equivalence class. Causal equivalence classes · 1 route
- Local intervention calculus when the DAG is absent (IDA) (opens background reference in a new tab)Estimate possible local causal effects without enumerating every compatible graph. Causal equivalence classes · 1 route
- Joint IDA with recursive regressions for causal effects (RRC) (opens background reference in a new tab)Estimate joint intervention effects across structures represented by an equivalence class. Causal equivalence classes · 1 route
Inspect the handoff
Data in; typed artifacts, diagnostics, and model-ready outputs out.
Each learning route declares its input shape, family-native artifacts, supported outputs, modeling assumptions, resource limits, and explicit refusals. The result stays typed so it can be inspected, exported, or handed to a compatible Darkstar model workflow.