AI Glossary / Causal inference and experimentation / Intermediate

Causal network

Definition

A causal network is a model that represents variables and their causal relationships, commonly using a directed graph together with probability distributions or structural mechanisms. In a causal Bayesian network, the model is intended to describe how outcomes respond to interventions, not only how observations are associated.

What it means

A conventional Bayesian network can compactly represent a joint probability distribution. Giving it a causal interpretation adds assumptions about the meaning of its arrows and what remains unchanged when a variable is deliberately set.
For example, observing that a machine alarm is active may raise the probability of a fault. Forcing the alarm indicator to display “active” does not necessarily cause the fault. A causal network can distinguish those questions if its mechanisms are correctly specified.
The network may be useful for explaining scenarios and planning interventions. Its answers depend on the validity of the causal structure, parameters and intervention assumptions, rather than on the visual clarity of the diagram.

How it works

Specify variables, their directed relationships and local mechanisms or conditional distributions. In a common acyclic model, the joint observational distribution factorizes into one conditional distribution per variable given its parents.
For an idealized intervention setting X to a fixed value, replace the mechanism that normally determines X with that assignment. The other mechanisms are assumed to remain unchanged. Compute the resulting outcome distribution under this modified model. This is different from simply conditioning on X=x in the original network. Real interventions may change several mechanisms or have side effects, so their model representation must match what will actually be done.

Three practical examples

• A reliability analyst compares repairing a component with merely observing that its warning indicator has changed.
• A process team explores how changing a controllable setting might affect quality through intermediate variables.
• A researcher uses a causal Bayesian network to examine which interventions could distinguish competing explanations of observed data.

Think of it this way

Think of a diagram of a machine's mechanisms rather than a chart of readings that tend to move together. Observing a dial tells you something about the machine; physically changing a mechanism can alter what happens next. The analogy explains observation versus intervention, but a causal network is only as credible as the mechanisms and assumptions encoded in it.

Worked example

Consider a simple model with a fault F that occurs with probability 0.10 and an alarm A. The alarm activates with probability 0.90 when a fault exists and with probability 0.10 when no fault exists. The graph is F→A.
Observing A=1 gives P(F=1|A=1)=(0.90×0.10)/(0.90×0.10+0.10×0.90)=0.50. The alarm therefore provides evidence about the fault.
Now imagine an idealized intervention that sets only the alarm display to active, without affecting the fault mechanism. Under that model, P(F=1|do(A=1)) remains 0.10. Conditioning and intervening answer different questions. The calculation assumes the display change truly leaves the fault process untouched; a real control that also alters equipment operation would require a different intervention model.

Illustrative case study

Fictional case: Sequoia Systems builds a network linking equipment condition, alarm readings and service interruptions. A dashboard user asks whether suppressing alarms would reduce interruptions because the two are associated.
The analyst explains that alarm suppression changes an indicator, while repairing the underlying condition changes a causal mechanism. The team reviews the graph with engineers and tests predictions using controlled maintenance records where feasible. It marks uncertain relationships and alternative explanations. This case illustrates how causal semantics can prevent a mistaken intervention argument; it does not establish that the network accurately represents every failure mode.

Common misconceptions and mistakes

A Bayesian network is not automatically causal. Observationally equivalent graphs can imply different intervention effects. Good prediction of observed data does not by itself validate the causal direction of every arrow.
Do not treat setting a recorded value as equivalent to changing the real mechanism it measures. Hidden common causes can alter what is identifiable. An acyclic representation may need time-indexed variables to represent feedback. Finally, a causal network can support intervention calculations without identifying every individual-level counterfactual question; stronger structural assumptions may be needed for those queries.

Business use

Causal networks can help organizations reason about interventions in complex systems, especially when domain knowledge and uncertainty must be combined explicitly. They can support diagnostic discussions, scenario analysis and decisions about additional experiments.
Use them to expose assumptions rather than to hide uncertainty behind a polished diagram. Document the source of each mechanism, the interpretation of intervention controls and the limits of available evidence. Validate observational predictions and intervention predictions separately where possible. If an operational change affects several processes at once, model those changes instead of representing it as a simple switch on one convenient variable.

People also ask

Is a causal network the same as a neural network?

No. It represents causal variables and mechanisms, although neural methods can sometimes be used within such models.

Can its arrows be learned from data alone?

Some structures can be investigated under additional assumptions, but association alone generally does not settle all directions.

Does observing X equal intervening on X?

Not in general.

Check your understanding

An alarm is caused by a fault. If a technician manually turns on the alarm indicator without changing the machine, should the model infer that the intervention caused a fault?

Answer: No, under the stated causal model. The intervention changes the indicator mechanism, while the fault process remains unchanged.

Related terms

• Causal graphs
• Bayesian network
• Structural causal model
• Conditional probability
• Confounding
• Causal graph
• intervention
• do-operator
• counterfactual

Learn next

Learn the difference between conditioning and intervention next. Use a two-variable cause-and-indicator example to compute both probabilities. Then examine which mechanisms a real proposed action would alter and whether the network contains enough information to predict that intervention.

More in Causal inference and experimentation

• Causal inference - Causal inference is the process of drawing conclusions about how outcomes would change under specified interventions.
• Causal language modeling - Causal language modeling is a training task in which a language model predicts the next token using only the preceding context.
• Causal link - In partial-order AI planning, a causal link records that one action establishes a condition needed by another action.
• Causal positivity - Causal positivity is the requirement that the treatment choices needed for a causal comparison have positive probability within the relevant covariate...
• Causality - Causality concerns relationships in which changes in one part of a system can produce changes in another.
• Causation - Causation is a cause-and-effect relationship in which one event, condition or intervention contributes to producing another.

Sources

Judea Pearl, Causal Diagrams for Empirical Research: https://bayes.cs.ucla.edu/R218-B.pdf bnlearn, Interventions and counterfactuals: https://bnlearn.com/examples/interventions/ AIMA worklist source: https://github.com/aimacode/aima-glossary/blob/master/glossary.md Primary-paper and official implementation excerpts consulted; full source-alignment and methodological review remain pending. The alarm example is original.

Previous: Causal link Next: Causal positivity