Is a Bayesian network the only graphical model?

No. Undirected models and factor graphs provide other ways to represent probabilistic structure.

Can it reason about hidden variables?

Yes. Inference can sum over hidden variables or estimate their distributions given observations.

Does a sparse graph always make inference easy?

No. Computational difficulty depends on structural properties, variable types, and the query, not just the number of visible edges.

AI Glossary / Graphs and symbolic AI / Intermediate

Probabilistic Graphical Model

Definition

A probabilistic graphical model represents a probability distribution using a graph and associated numerical functions. The graph helps express how variables are organized and which conditional independence assumptions the model makes. Bayesian networks, Markov random fields, and factor graphs are major forms. The model supports reasoning about uncertain quantities, including quantities that have not been observed directly.

What it means

A full probability table over many variables can become enormous. Graphical models use structure to represent a distribution through smaller pieces when the chosen assumptions permit it. The graph makes those assumptions inspectable rather than hiding them inside one undifferentiated table.
The graph is not the whole model. Its variables need clear meanings, and the associated probabilities or factors need values. Nor is every diagram of connected concepts a probabilistic graphical model. A knowledge graph, flowchart, or neural-network drawing may serve a different purpose unless a corresponding probabilistic interpretation is explicitly defined.

How it works

In a Bayesian network, a directed acyclic graph associates each variable with a conditional distribution given its parents. Multiplying those local distributions gives the joint distribution. An undirected model instead combines compatible factors over groups of variables and generally needs a normalization constant. Factor graphs show variables and factors explicitly as different node types.
Inference uses the resulting model to answer queries, often conditioning on observations and summing over hidden variables. Learning estimates parameters or structure from data. These are related but distinct tasks. A compact representation can reduce storage yet still make some inference queries computationally demanding, depending on the graph and the requested calculation.

Three practical examples

A maintenance-analysis model can connect equipment state, sensor readings, and operating conditions to support uncertainty-aware inspection.
A language-processing model can represent dependencies among hidden sequence states and observed symbols.
A service-planning simulation can express relationships among workload, delays, and observed status messages while keeping its assumptions open to review.

Think of it this way

Think of a map that shows which local instruction sheets must be combined to describe a whole system. The map organizes the information, while the sheets supply the numerical details. The analogy is limited because graphical models encode specific probability factorizations and independence properties, not simply proximity or informal relationships.

Worked example

Consider an invented Bayesian network with binary variables F for a machine fault, A for an alarm, and C for an inspection call. Use arrows F → A → C and factorization P(F,A,C) = P(F)P(A|F)P(C|A).
Set P(F) = 0.1, P(A|F) = 0.9, and P(A|not F) = 0.2. Set P(C|A) = 0.8 and P(C|not A) = 0.1. The joint probability of fault, alarm, and call is 0.1 × 0.9 × 0.8 = 0.072.
The alarm probability is 0.1 × 0.9 + 0.9 × 0.2 = 0.27. Therefore the call probability is 0.27 × 0.8 + 0.73 × 0.1 = 0.289. This calculation sums over unobserved possibilities rather than selecting one most likely state.
The graph assumes that, once the alarm state is known, the fault provides no additional information about the call: P(C|A,F) = P(C|A). That is a modeling assumption to examine, not a fact established by drawing the arrows. If staff also call after directly noticing a fault without an alarm, this simple structure may be inadequate.

Illustrative case study

In a fictional factory analytics project, a team draws a network linking machine condition, alarm messages, and operator calls. Early validation finds that operators sometimes call because of a visual observation that the model omitted. The apparent alarm-call relationship changes across shifts with different inspection routines.
The team adds a clearer observation model and reviews the relevant conditional assumptions before fitting new parameters. It checks predictions on held-out operating periods and examines calibration. The case illustrates model revision driven by evidence, not a claim that a graphical representation alone creates reliable operational insight.

Common misconceptions and mistakes

An arrow in a Bayesian network is not automatically proof of causation. A causal interpretation requires additional assumptions about how the system works and what interventions mean. A network fitted to observational data may represent a distribution without answering intervention questions.
Also, a visible path does not guarantee dependence for every possible choice of parameters. Graph separation provides certain independence guarantees, while particular parameter values can introduce additional independencies. Finally, conditional independence is not the same as marginal independence: learning another variable's value can remove or introduce associations relevant to inference.

Business use

The business value is a structured account of uncertainty that can be inspected, challenged, and updated. Stakeholders can discuss missing variables and independence assumptions before relying on a numerical prediction. The same structure can support several queries without building a separate ad hoc rule for each.
Validation still needs representative data, credible parameter estimates, and a clear intended use. Record what is observed, what is hidden, and how missing or noisy evidence is handled. Compare model complexity with the amount of available data and the inference budget. Treat the model's probabilities as claims requiring evaluation, not as guarantees produced by the graph.

Check your understanding

In the worked model, what is the probability of no call?

Answer: It is 1 − 0.289 = 0.711. Equivalently, 0.27 × 0.2 + 0.73 × 0.9 = 0.054 + 0.657 = 0.711.

Related terms

Learn next

Next, study conditioning and marginalization in the worked network. Compute P(F|A) as 0.09/0.27 = 1/3 and explain why the alarm raises the fault probability without making a fault certain. Then identify which modeling assumptions would change if operators could call after a separate direct inspection.

Sources

Stanford CS228 course notes, Bayesian networks and Markov random fields: https://ermongroup.github.io/cs228-notes/representation/directed/ and https://ermongroup.github.io/cs228-notes/representation/undirected/ . The original course page links these notes: https://ai.stanford.edu/~ermon/cs228/index.html . Relevant representation sections were read. The machine example is original.

Previous: Probabilistic forecasting Next: Probabilistic Horn abduction