Description
Represent many interrelated discrete variables as a directed graph with probability tables, fill the tables from data and expert judgement, and update beliefs about any variable when others are observed.
A Bayesian network is a directed acyclic graph whose nodes carry conditional probability tables given their parents; multiplying the tables gives the joint distribution, and d-separation reads off which variables are independent given others. Tables are learned by counting cases with Dirichlet priors, which keep unseen combinations above zero, and can be blended with expert estimates and updated as data arrive. Evidence may be certain, a revised distribution for a node (Jeffrey's rule) or an uncertain report expressed as likelihood ratios (Pearl's virtual evidence). Queries such as updated marginals or the most probable explanation are answered by variable elimination or junction tree message passing, in either direction: from causes to effects or from symptoms back to likely causes. Sensitivity analysis shows which table entries would flip a conclusion.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Siloed
- Stage
- Discovery, Evaluation, Post-launch
- Data
- Mixed
- Effort
- Weeks
Steps
- Choose the variables and a small set of discrete states for each.
- Draw the graph with arrows from each variable to those it directly influences, from domain knowledge or a structure-learning algorithm, and check the implied independences with d-separation.
- Attach a conditional probability table to each node given its parents, keeping parent sets small because tables grow exponentially.
- Fill the tables from data with Dirichlet priors, or from expert estimates where data are thin, and update them as cases arrive.
- Enter evidence as certain findings, soft findings or likelihood ratios and compute the updated probabilities or most probable explanation.
- Test the sensitivity of key queries to table entries and revise the network where small changes flip the conclusion.
Sources
- Bayesian Networks: An Introduction (Koski & Noble, 2009)
Also known as
- Bayesian network
- Bayesian belief network
- Belief network
- BN
- Directed probabilistic graphical model
- Junction tree inference
- Evidence propagation
Categories
- Analytics & Behaviour
- Statistics & Modelling