Must a chain graph contain both edge types?

No. Directed acyclic graphs and undirected graphs are included as special cases.

Is a chain component always one variable?

No. It can contain multiple variables joined by undirected connections.

Are LWF and AMP interchangeable?

No. They agree in important special cases but can assign different conditional-independence meanings to mixed graphs, so the intended interpretation should be stated.

AI Glossary / Probabilistic machine learning / Intermediate

Chain graphical model

Definition

A chain graphical model represents a probability distribution using a graph that can contain both directed and undirected edges, subject to restrictions that prevent partially directed cycles. It combines features of directed and undirected graphical models. This entry uses the mixed chain-graph meaning, rather than a graph that simply forms a linear sequence.

What it means

Some modeling problems naturally separate variables into groups. Relationships within a group may be represented without choosing a direction, while relationships between groups have an ordering.
A chain graph accommodates this combination. Its undirected connected groups are called chain components, and arrows between those components must form an acyclic structure. The components need not lie along one simple line: the word chain does not require a single sequence of nodes.
The graph's meaning also depends on its Markov interpretation. LWF and AMP are established interpretations with different conditional-independence semantics for some mixed structures. Naming the graph alone is therefore not always enough to specify the probabilistic model.

How it works

Choose variables and identify the groups whose internal relationships will be represented by undirected edges. Draw arrows between groups according to the intended ordering, checking that no cycle can be followed consistently through arrow directions and undirected links while traversing at least one arrow.
Specify the interpretation and parameters. In a common LWF construction, the joint distribution factors into a product of conditional distributions for chain components given their parent variables. Each component conditional has an appropriate undirected factorization and normalization.
Use the resulting distribution for estimation or inference. Conditional-independence claims must follow the selected interpretation rather than a casual mixture of Bayesian-network and undirected-graph rules.
An arrow does not by itself establish causation. A causal interpretation requires additional assumptions about the system and how interventions relate to the model.

Three practical examples

An analyst models two jointly related measurements within each processing stage while representing dependencies from earlier stages with arrows.
A researcher uses a conditional random field to model related output labels given inputs, connecting this conditional structure to a mixed-graph representation.
A modeling team compares a fully directed specification with one that preserves undirected relationships inside a group of contemporaneous outcomes.

Think of it this way

Think of an organization diagram with ordered project stages and teams that collaborate within each stage. The diagram can distinguish progression between stages from relationships inside a team without assigning every internal relationship a direction.
The analogy has limits: probability graphs encode formal constraints on distributions, not just organizational descriptions. A visual grouping is useful only when its mathematical interpretation and assumptions are specified.

Worked example

Consider an original three-variable example with binary A, B, and C. Draw A → B, A → C, and B - C. The chain components are {A} and {B,C}; the directed ordering goes from the first component to the second.
Specify p(A=1) = 0.4. Given A=1, let the probabilities of (B,C) = (0,0), (0,1), (1,0), and (1,1) be 0.1, 0.2, 0.2, and 0.5. They sum to one. Given A=0, use 0.5, 0.2, 0.2, and 0.1, also summing to one.
The factorization is p(A,B,C) = p(A) × p(B,C | A). Thus p(A=1,B=1,C=1) = 0.4 × 0.5 = 0.2.
Within the A=1 group, p(B=1 | A=1) = 0.7 and p(C=1 | A=1) = 0.7. Their product is 0.49, which differs from the joint probability 0.5. The component conditional therefore should not be replaced by independent B and C factors. This deliberately simple graph illustrates grouping without attempting to distinguish all LWF and AMP cases.

Illustrative case study

In a fictional project, Rowan Manufacturing models measurements from two stages of a process. The team initially directs every connection, although it has no substantive reason to order two measurements taken within the same stage.
It examines a chain-graph specification and explicitly records which interpretation and conditional factors it uses. The analysts check its assumptions against the data and avoid calling the internal undirected link a proven causal relationship.
This invented case illustrates a modeling choice. It reports no demonstrated performance advantage for chain graphs over other models.

Common misconceptions and mistakes

Do not confuse a mixed chain graph with a Markov chain, which typically describes a sequence with a particular dependence structure.
Do not assume that every graph mixing arrows and undirected edges is a valid chain graph. Partially directed cycles are excluded.
Do not read all missing edges with one universal rule: the interpretation matters.
Do not assume that an undirected link represents two opposite causal arrows.
Finally, do not replace a joint component conditional with independent node conditionals unless the specified distribution supports that factorization. Doing so can remove dependencies the model was intended to represent.

Business use

Chain graphical models can help teams express dependencies when a fully directed or fully undirected description is too restrictive for their intended assumptions. They offer a structured language for discussing groups of related variables and dependencies between groups.
The practical cost is additional care in specification and interpretation. Record the chain components, the chosen Markov semantics, and the actual factors used for computation. A clear graph is useful for review, but model quality still depends on the data, the parameterization, and whether the encoded assumptions fit the problem.

Check your understanding

In the worked example, why does p(B=1,C=1 | A=1) not equal the product of the two conditional marginals?

Answer: The specified joint component distribution retains dependence between B and C. Its joint value is 0.5, while multiplying the marginals gives 0.49.

Related terms

Learn next

identify chain components in small mixed graphs and check for forbidden partially directed cycles. Then compare LWF and AMP interpretations on an example where their conditional-independence implications differ.

Conditional independence

Sources

Primary teaching source: Ravikumar, Chain Graphical Models, Carnegie Mellon 10-708 notes: https://www.cs.cmu.edu/~pradeepr/courses/708/2020-fall/resources/Markov-CGMs.pdf The graph, probability tables, and fictional case are original illustrations. The supplied PML book PDF was not accessible for checking its precise use of this term.

Previous: Certification Next: Chain linking