Open-source agentic AI for transaction monitoring. Just ask Jooby.What's Jube? →
jube
  • Documentation
  • Analytical Methodology
  • Bayesian Networks
  • Explanatory Role Inside AdvancedAnalytics

Explanatory Role Inside AdvancedAnalytics

What is real and callable today

One Bayesian-network-shaped capability ships in the product: Tree-Augmented Naive Bayes (TAN). Training runs through Jube.Service.Repository.AnalyticsTrainingRun.AdvancedAnalyticsTreeAugmentedNaiveBayesClassifierTrainingService, invoked as an agent tool by Jube.Service.Agent.AdvancedAnalyticsTools.TreeAugmentedNaiveBayesClassifierTrainingRunTools / "SubmitTreeAugmentedNaiveBayesClassifierTraining" (BayesianAnalysisWithBnlearn.md, Procedure 7). It genuinely runs the same Chow-Liu maximum-weight-spanning-tree algorithm over class-conditional mutual information that tree.bayes runs in this subsection, fits a Gaussian per predictor conditioned on the class and, where a predictor has one, its tree parent, and reports a calibrated probability.

CompareModels on Jube.Service.Agent.AdvancedAnalyticsTools.ModelEvaluation.ModelSelectionTools, with EvaluationFamily.TreeAugmentedNaiveBayes selected, cross-validates that classifier against up to eleven other families on identical folds. That is the whole of what is shipped: one class variable, every other variable a predictor beneath it, one designated tree structure discovered once per training run, one forward probability per scored row.

What is not

Everything else worked through in the earlier files of this section — graphical model foundations, constraint-based, score-based and hybrid structure learning, parameter learning, inference, and network validation — is bnlearn-only, with no product path:

  • Arbitrary structure, hand-drawn (model2network) or learned by general-purpose search (hc, tabu) — any DAG shape over any set of nodes, not one fixed class-plus-predictors tree. Out of scope per docs/AnalyticalMethodology/Reference/ModelCatalogue.md: “General Bayesian network with structure search… Drop: KDB, TAN, AODE and HNB give most of the benefit.”

  • Hand-set or learned conditional probability tables over more than a classifier-shaped tree — custom.fit for subjective belief tables with no data behind them, or bn.fit fitting an arbitrary structure’s tables from data, both covered in BayesianAnalysisWithBnlearn.md. TAN’s training service fits its own tables automatically as part of one fixed procedure; there is no tool that fits or accepts a table for a standalone node outside that procedure.

  • Multi-node querying — cpquery conditioning on any node and reading off any other node’s posterior, in either direction (BayesianAnalysisWithBnlearn.md, Procedure 2’s forward-prediction-and-explanation pair). A trained classifier, TAN included, only ever answers the one forward question its training fixed: the class variable’s probability given a full row of predictor values. There is no notion of picking a different node as the query target, because there are no other queryable nodes — the predictors are terminal inputs, not addressable points in a graph.

  • Intervention — mutilated() and the conditioning-versus-do() distinction worked through in NetworkValidationAveragingAndCausality.md. Nothing in AdvancedAnalyticsTools deletes an arc, fixes a node and re-queries the rest of a graph, because nothing in AdvancedAnalyticsTools has arcs to delete.

None of this is an oversight so much as a considered scope boundary: the catalogue’s own reasoning for excluding general structure search is that the classification benefit is already covered more cheaply by KDB, TAN, AODE and HNB-style methods, not that a general network was overlooked.

How this actually runs today, outside the classifier path

“No product path” above means no native [AdvancedAnalyticsTool] — it does not mean there is no way to actually run any of this. The R script execution job pattern (Jube.R, the sandboxed R server, and the RScriptTools submit/poll/ result agent tools; see R Server) lets an agent submit a genuine bnlearn script as a job and read back its console output: arbitrary structure (model2network, hc, tabu), hand-set or learned CPTs (custom.fit, bn.fit), multi-node querying in either direction (cpquery), and intervention (mutilated()) are all real, callable operations this way — everything this file lists as “bnlearn-only, with no product path” is bnlearn -only in the sense of having no native tool, not in the sense of being unreachable.

This is the mechanism behind Governance, KPIs and Model Validation, which names Bayesian networks “the primary probabilistic framework for both anomaly detection and prescriptive analytics” and the specific fallback methodology for weak or absent labels — unsupervised, probabilistic detection via conditional probabilities over feature states, calibrated indirectly (posterior-distribution stability, anomaly-score consistency over time, expert review of high-risk tails) exactly where label-based calibration cannot be used. That is deliberately a batch, expert-reviewed use, run as a job and read back, not a per-transaction score — there is no live recall path, which this fallback methodology does not call for.

Why the material still earns its place

A TAN classifier is a Bayesian network — a small, fixed-shape one, but a real one, with real conditional independence assumptions baked into its structure. Reasoning about what it is actually saying requires the same vocabulary this subsection built on general networks, only applied to the one shape that exists in product:

  • Conditional independence. TAN’s entire reason for existing over plain naive Bayes is a conditional independence claim: every predictor is independent of every other predictor given the class, except for one chosen parent predictor per child. Reading a TAN model’s reported per-predictor evidence (BayesianAnalysisWithBnlearn.md, Procedure 7) without understanding that assumption invites over-reading a predictor’s isolated contribution as if it captured every interaction it participates in, when by construction it only captures its single strongest pairwise one.

  • Markov blankets. In any Bayesian network, a node’s Markov blanket — its parents, its children and its children’s other parents — is the exact set of other nodes that renders it conditionally independent of everything outside that set. In TAN’s tree, the class variable’s Markov blanket is every predictor (since every predictor is a child of the class); a predictor’s Markov blanket is the class, its own tree parent (if it has one) and its tree children. That is a precise, checkable statement about which of a TAN model’s reported predictor contributions can shift if a different, currently unreported variable changes, and which cannot — the same reasoning shd() and the true asia DAG in NetworkValidationAveragingAndCausality.md depend on to say a learned structure is “close” to a real one at all.

  • Structure learning as a claim, not a fact. NetworkValidationAveragingAndCausality.md’s cross-validation and bootstrap material makes the point that a learned structure is one noisy estimate, not ground truth — arc presence and direction both carry a confidence, and a single hill-climbing run can get an arc’s direction wrong even when the arc itself is well supported. TAN’s Chow-Liu tree is fitted once, deterministically, from one seeded split, with no bootstrap or cross-validated-structure step of its own (BayesianAnalysisWithBnlearn.md, Procedure 7’s implementation notes; see also this file’s “what is not” list). That does not make its parent-selection wrong, but it means the same caution applies: the reported tree is what the Chow-Liu algorithm judged the maximum-weight spanning structure over one sample, not a verified causal fact about which predictor genuinely drives another.

  • Calibrated uncertainty. TAN’s raw posterior, like plain naive Bayes, is typically overconfident, which is exactly why its training service fits a Platt calibration curve on a held-out split before reporting a probability (BayesianAnalysisWithBnlearn.md, Procedure 7’s tool description). Recognising why an uncalibrated Bayesian posterior overstates its own confidence — because a tree-structured product of conditional likelihoods compounds every individual conditional’s own error — is general Bayesian-network reasoning, not something specific to TAN, and it is the same reasoning that explains why cpquery’s conditioning answers in NetworkValidationAveragingAndCausality.md needed 2,000,000 weighted-logic-sampling draws to stabilise rather than being read off a table directly.

  • Conditioning versus intervention. This is the sharpest transfer. A TAN model’s predictors are genuinely just conditioning evidence: scoring a row conditions on every predictor’s observed value and reads off the class posterior, with no claim anywhere that changing one predictor’s value in the real world would move the class probability the way the model’s own sensitivity to that predictor might suggest.

    NetworkValidationAveragingAndCausality.md’s kidney-stone example is not a TAN-specific concern, but it is exactly the trap an analyst falls into when they read a strong predictor-to-class relationship out of any classifier’s evidence — TAN, logistic regression, a random forest’s feature importance, all of it — and describe it to a decision-maker as “if we changed this, the risk score would fall,” which is an interventional claim a conditional model was never built to support. Knowing to ask “is there a plausible common cause behind both this predictor and the outcome” before making that kind of statement, and knowing that answering it properly would need something like a mutilated graph rather than another conditional query, is the concrete, transferable payoff of working through NetworkValidationAveragingAndCausality.md on a toy network first.

That vocabulary is not confined to TAN. Conditional independence, calibration and the conditioning-versus- intervention distinction apply equally to general statistical reasoning elsewhere in AdvancedAnalyticsTools — reading a hypothesis test’s p-value, a regression coefficient, or a clustering’s silhouette score correctly all benefit from the same discipline about what was actually conditioned on versus what was actually varied.

The bounded claim

“Explanatory power” here means specifically: the concepts in this subsection transfer to reasoning carefully about the one Bayesian-shaped model the product ships, and to general statistical reasoning across AdvancedAnalyticsTools more broadly. It does not mean the product can act on an arbitrary learned Bayesian network, answer a multi-node query, or compute an interventional probability — those remain bnlearn-only, demonstrated here for the understanding they build, not for a recall path that does not exist. An analyst using this material to describe a candidate causal story to a human decision-maker should say what the data conditionally supports and be explicit about what would be needed to say more (an intervention, a designed experiment, or a domain-known absence of confounding) rather than reporting a conditional relationship as if it already answered the causal question.