Part IV · Ch. 19 — Bayesian Network

draft

Probabilistic · Causal/diagnostic inference · Directed acyclic graph (DAG). Represents conditional dependencies between variables and infers posterior probabilities. Card: ../02-types-of-ai/03-probabilistic.kmd.

🎨 Figure F-IV.19.0Belief that propagates. Brief: DAG of variables (e.g., symptom→disease) with probability tables; observed evidence updating posteriors across the graph.

Bayesian network (DAG)

1. Definition and short history

Probabilistic graphical model (Judea Pearl, 1980s-90s; Turing Award 2011) that factors a joint distribution by dependencies. Foundation of causal reasoning.

2. Foundations

  • Probability — Bayes' rule, conditional independence.
  • Graph theory — DAG, d-separation.
  • Causal inferencedo-calculus, interventions.
  • Statistics — parameter estimation.

3. Algorithms and architectures

  • Structure — DAG + conditional probability tables (CPTs).
  • Exact inferencevariable elimination, junction tree.
  • Approximate inferencebelief propagation, MCMC, variational.
  • Learning — of structure and of parameters (from data).

4. Inputs

  • Hardware: CPU; inference can be expensive (NP-hard in general).
  • Data: observations to estimate CPTs; domain knowledge for structure.
  • Data structures: DAG, CPTs, factor graphs.
  • Systems: pgmpy, probabilistic inference libraries.

5. Specialized lifecycle

Stage Specialization
0 Problem Diagnosis/risk with uncertainty and need for causal explanation
1 Data Observations + structure knowledge
2 EDA Correlations, independences, missing data
3 Modeling Define/learn DAG; CPTs
4 Training Estimate parameters (and structure) from data
5 Evaluation Calibration, log-likelihood, diagnostic accuracy
5.5 Acceptance Validate causal structure with experts
6 Production Posterior inference given evidence; explainable
7 Monitoring Distribution drift
8 Retraining Re-estimate with new data
9 Governance Causal transparency, use in sensitive decisions

6. Capabilities, modes and modalities

Intellectual/diagnostic: medical diagnosis, risk assessment, causal reasoning, sensor fusion; calibrated and explainable uncertainty.

7. Limits, risks and ethics

Exact inference is expensive; causal structure is hard to learn; does not scale to unstructured data. Strength: causality and interpretability.

8. State of the art and examples

Diagnosis, risk, causal inference (Pearl); convergence with deep learning in deep probabilistic models and causal ML; relevant where cause matters more than correlation.