What is a Bayesian network?
A Bayesian network is a graphical model of uncertainty. Each variable in a problem, such as “pump bearing worn” or “vibration high,” is a node. Arrows connect causes to effects, and the graph may not contain cycles. Each node carries a table giving the probability of its values for every combination of its parents’ values. Together the graph and the tables define the joint probability of every variable at once, in a form people can read and experts can edit.
Unlike most machine learning models, a Bayesian network can be built from expert knowledge, learned from data, or both. That makes it a bridge between knowledge-driven AI and data-driven AI, and one of the most interpretable probabilistic tools available.
How a Bayesian network works
The model is used by entering evidence and asking questions. Set “vibration high” to true and “oil temperature” to its observed value, and the network recomputes the probability of every unobserved node, including the likely causes. This inference runs in either direction along the arrows: from causes to predicted effects, or from observed symptoms back to probable causes. The probability tables can be estimated from historical records by counting. When the structure itself is unknown, structure learning algorithms search for the graph that best explains the data, constrained by what experts know cannot be true.
Bayesian networks in practice
A paper mill suffers intermittent sheet breaks. Engineers and an analyst build a Bayesian network with nodes for moisture, machine speed, pulp consistency, roll wear, and the break event, with probabilities estimated from two years of process data and adjusted by the operators. When a break occurs, the network ranks the likely causes given the current readings, which shortens diagnosis. A hospital uses a similar model for triage, combining symptoms, vital signs, and history into probabilities for several conditions.
What to watch for
- Arrows express conditional dependence, and a learned arrow is not proof of causation unless the structure was built with causal knowledge.
- Tables grow quickly with the number of parents per node, so keep the graph sparse and discretize continuous variables with care.
- Expert-supplied probabilities should be checked against data where any exists, since confident experts are often miscalibrated.
- Bayesian networks suit problems with tens to hundreds of variables, while raw images or text call for neural networks.