Does a root node have a CPT?

Its distribution has no parents, so it can be viewed as an unconditional table.

Can a CPT contain zeros and ones?

Yes, when the model assigns impossible or certain outcomes, but those choices need justification.

Must Bayesian networks use explicit tables?

No. Conditional distributions can have other representations, especially for continuous variables or structured dependencies.

AI Glossary / Statistics and probability / Intermediate

Conditional probability table

Definition

A conditional probability table, or CPT, is a tabular representation of a discrete variable's probability distribution for each specified combination of conditioning-variable values. In a Bayesian network, a node's CPT typically gives its probabilities conditional on its parent nodes. For every fixed parent configuration, probabilities across the node's possible states must sum to one. The table encodes probabilistic relationships, not automatically causal ones.

What it means

A CPT answers questions such as: given these operating conditions, how likely is each equipment state? It contains a separate distribution for each condition combination. A binary outcome can be represented by one probability per combination because the other probability is its complement, although software may require both explicitly.
A CPT differs from a joint probability table. Its entries are normalized separately within each conditioning configuration, rather than all entries together summing to one. The orientation of rows and columns varies between books and software, so labels matter more than visual layout.

How it works

Define the target variable, its mutually exclusive and exhaustive states, and the variables being conditioned on. Enumerate their possible configurations. Assign probabilities using suitable data, expert elicitation, or a specified model, documenting how the values were obtained.
Check each distribution for nonnegative entries that sum to one. Confirm state order and parent order. A correct set of numbers attached to the wrong labels can produce a mathematically valid but substantively incorrect model.
If the target has r states and the conditioning variables have q possible configurations, a full table contains rq probabilities and q(r - 1) freely varying parameters after normalization. For k binary parents, q=2 to the power k. The table can therefore grow quickly as parents are added.
In a Bayesian network, the graph and all local conditional distributions define a joint distribution through multiplication of each node's probability given its parents. Inference can then combine evidence and sum over unobserved variables. A CPT for one node alone does not supply the frequencies of its parent configurations. Learning large tables from limited data may require smoothing, structured representations, or a simpler model.

Three practical examples

A diagnostic model stores equipment-failure probabilities for combinations of load category and maintenance status.
A document model represents category probabilities conditional on discrete indicators extracted from a message.
A planning model records possible delivery outcomes for combinations of route conditions and service types, then combines them with a model of how often those conditions occur.

Think of it this way

Think of a CPT as a collection of labeled probability cards. Each card corresponds to one combination of known conditions and contains a complete distribution over the outcome. Choosing the right card is essential. To average across cards, you also need to know how frequently their conditions occur.

Worked example

Consider an original CPT for a binary disruption D, with binary parents high load H and overdue maintenance M. Define the probabilities P(D=yes given H,M) as follows: neither condition, 0.02; high load only, 0.10; overdue maintenance only, 0.15; both, 0.40. The corresponding no-disruption probabilities are 0.98, 0.90, 0.85, and 0.60.
There are four parent configurations and two outcome states, so the full table has eight entries but only four free probabilities. Each yes/no pair sums to one; the entire table sums to four, which is appropriate because it contains four separate distributions.
Suppose P(H=yes)=0.30 and P(M=yes)=0.20, and explicitly assume H and M are independent for this example. The configuration probabilities are 0.56, 0.24, 0.14, and 0.06 in the order above. The marginal disruption probability is 0.56 times 0.02 + 0.24 times 0.10 + 0.14 times 0.15 + 0.06 times 0.40 = 0.0802.
Without the parent probabilities and their dependence assumptions, the CPT alone would not determine that 8.02% marginal probability. These invented values illustrate inference, not equipment reliability data.

Illustrative case study

In a fictional company, Harbor Equipment builds a small Bayesian network for operational planning. Engineers supply an initial CPT for disruption under different operating conditions, and analysts compare those judgments with available records.
A review finds that two software columns use a different maintenance-state order from the engineers' worksheet. The team corrects the labels and checks every configuration before interpreting any output. It also identifies rare combinations whose probability estimates are poorly supported.
Managers receive scenario results with the assumptions and uncertain entries clearly identified. This fictional case illustrates the importance of labeling, validation, and evidence; it does not claim that the model provides a reliable causal explanation of breakdowns.

Common misconceptions and mistakes

Do not normalize the entire CPT as if it were one joint distribution. Do not reverse the conditioning direction by reading the table backward. A network arrow does not automatically establish causation. An observed zero count need not mean an event is impossible. Adding parents can rapidly increase data requirements. Independent root variables in a particular network are a model assumption, not a consequence of writing a table with two parent columns.

Business use

Businesses use CPTs in interpretable probabilistic models for diagnosis, planning, and structured risk assessment. Their explicit entries make assumptions inspectable, but inspectability does not guarantee accuracy. Record the data or judgment behind each configuration, validate predictions, and examine how uncertain entries affect decisions. For large or continuous feature spaces, a full discrete table may be impractical or lose information through coarse categorization.

Check your understanding

How many free parameters does a three-state target with four parent configurations need in an unrestricted CPT?

Answer: Four times two equals eight.

Why does the example's complete table sum to four?

Answer: It contains four separately normalized distributions.

What extra information is needed to calculate an overall disruption rate?

Answer: The joint distribution of the parent configurations.

Related terms

Learn next

Learn conditional probability and Bayesian-network factorization, then construct a small table by hand. Verify state ordering and normalization before calculating marginal probabilities. Next explore inference and parameter estimation, including how sparse observations and structured conditional distributions affect model complexity.

Sources

Poole and Mackworth, Artificial Intelligence: Foundations of Computational Agents, third edition, Belief Networks: https://www.artint.info/3e/html/ArtInt3e.Ch9.S3.html AIMA glossary, conditional probability table: https://github.com/aimacode/aima-glossary/blob/master/glossary.md

Previous: Conditional probability distribution Next: Conditional random field