How many conditioning cases does a node with no parents have?

One. Its distribution is unconditional within the network specification and can be represented by a single row.

Is a missing parent value the same as a particular case?

No. It leaves several cases possible, which inference must handle using the available evidence and model.

Must every case be equally likely?

No. The local table describes child probabilities given each case, not the probabilities of the cases themselves.

AI Glossary / Graphs and symbolic AI / Intermediate

Conditioning case

Definition

A conditioning case is one combination of values assigned to a Bayesian-network node's parents. In a conditional probability table, it identifies the parent setting for which a distribution over the child node is specified. Each case answers the same child-variable question under a different set of parent values.

What it means

Suppose a child variable has two parents. Knowing only one parent's value does not usually identify a complete conditioning case, because the other parent can still take several values.

The case is a local context for the child distribution, not an assignment to every network variable. Other variables may exist elsewhere without becoming part of this node's probability table.

For a discrete child, the probabilities of its possible values must sum to one within each conditioning case. Different cases describe different conditional distributions; their rows do not need to sum to one when added together. The table does not by itself specify how common each parent combination is.

How it works

List the possible values of every parent and form their combinations. Each combination identifies a row or equivalent slice of the conditional probability table. Table layout can vary, so labels are more reliable than assuming a particular visual orientation.

For parents with domain sizes k1, k2, and so on, the full tabular representation has their product as its number of conditioning cases. If the child has r values, each case contains r probabilities, of which r minus one are freely specified after normalization.

When all parent values are observed, select the matching case. When some are unknown, combine relevant cases using appropriate probabilities for those unknown parent values, given the available evidence. Selecting an arbitrary row or taking an unweighted average generally answers the wrong question.

Sparse data can leave some cases poorly estimated or unobserved. That is an estimation problem, not evidence that the missing case is impossible. The model needs an explicit treatment of such gaps.

Three practical examples

1. A workflow model gives a different distribution of processing outcomes for each combination of workload category and resource availability.

2. A Bayesian-network reviewer checks every parent combination to find omitted table entries and verify that each child distribution is normalized.

3. A model designer estimates how much a new discrete parent will enlarge a probability table before adding that dependency to the network.

Think of it this way

Think of a conditioning case as a labeled drawer containing the probabilities for one exact setting of the parent controls. Changing a parent value leads to a different drawer.

The analogy has limits: the drawer contains a probability distribution, not a predetermined outcome. Also, when a control setting is unknown, inference must combine relevant drawers with justified weights rather than pretending one has been selected.

Worked example

Let an outcome R have three values: clear, review, and defer. Its parents are mode M, with values A, B, and C, and availability S, with values ready and unavailable.

There are 3 times 2 = 6 conditioning cases. The table stores 6 times 3 = 18 probabilities. Because each case's three probabilities sum to one, there are 6 times 2 = 12 freely specifiable numbers.

For S = ready, suppose the distributions for modes A, B, and C are respectively (0.7, 0.2, 0.1), (0.4, 0.4, 0.2), and (0.1, 0.3, 0.6), in the stated outcome order. Knowing M = B and S = ready selects the second distribution, so the probability of review is 0.4.

If only S = ready is known, suppose the conditional mode probabilities are 0.5, 0.3, and 0.2. The review probability becomes 0.5 times 0.2 plus 0.3 times 0.4 plus 0.2 times 0.3 = 0.28.

The other mixed probabilities are 0.49 for clear and 0.23 for defer. Together they sum to one. An unweighted average would ignore the supplied mode probabilities.

Illustrative case study

Fictional case: Cedar Inference builds the six-case model above. An analyst treats ready as a complete case and copies the mode-A row into a report, although the mode is unknown.

The reviewer identifies the missing parent value and asks for the mode distribution conditional on readiness. The report is recalculated by combining the three relevant cases with the stated weights. The team also labels every row with both parent values to reduce ambiguity.

This case illustrates a local probability-table interpretation problem. It does not assert measured accuracy gains or claim that the invented outcome probabilities describe a real workflow.

Common misconceptions and mistakes

Do not confuse a conditioning case with a complete possible world for the entire network. It fixes the parents relevant to one local distribution.

Do not sum probabilities down an arbitrary column and expect one. Normalization applies across the child's alternatives within a fixed parent case.

Do not multiply parent domain sizes and then also count child values as additional conditioning cases. Child values are outcomes within each case. Finally, do not assume equal weights when some parents are unknown unless the model actually supports that assumption.

Business use

For business teams, conditioning cases make model detail and data requirements concrete. A node with many categorical parents can need a large number of local distributions, each requiring evidence or a defensible specification.

This encourages careful discussion of whether every dependency is necessary and whether rare parent combinations have adequate support. More compact parameterizations can sometimes replace a large table, but they introduce their own assumptions. Review should cover both the meaning of each parent setting and the reliability of the associated probabilities.

Check your understanding

A node has one Boolean parent and one parent with four values. How many conditioning cases does its full table contain?

Answer: Eight, because 2 times 4 = 8. The number of child values determines probabilities within each case, not the case count.

Related terms

Learn next

Construct a small table with parents of different domain sizes and label every case explicitly. Then practice answering queries with one parent unobserved, paying attention to the conditional weights needed to combine rows.

Sources

Russell and Norvig, Artificial Intelligence: A Modern Approach, author-hosted corrected page on Bayesian networks and conditional probability tables: https://aima.cs.berkeley.edu/errata/aima-414.pdf . The definition and normalization discussion were consulted. The table, calculation, and fictional review are original.

Previous: Conditioning Next: Confabulation (neural networks)