Why use collective inference?

Unknown attributes can inform one another through modeled relationships, rather than being predicted independently.

Are the edges causal?

Not by default. They represent predictive dependencies in the chosen model.

Can independently trained conditionals be inconsistent?

Yes. This is a recognized issue in dependency networks and affects how their probabilistic interpretation and inference should be understood.

AI Glossary / AI and machine learning / Advanced

Relational dependency network

Definition

A relational dependency network, or RDN, is a probabilistic graphical model that represents dependencies among attributes of linked objects using a collection of local conditional models. It supports reasoning about relational data in which an object's unknown attributes may depend on both its own information and information about related objects.

What it means

Some predictions are connected. The likely topic of one document may be informative about documents linked to it, while their likely topics can also inform the first prediction. An RDN provides a framework for modeling such dependencies without requiring the acyclic structure used by a Bayesian network. Its links express statistical dependence under the model, not necessarily cause and effect. The word relational refers to objects and their relationships, which can be represented through linked records or graph-like data.

How it works

Learn local conditional distributions for attributes using information from an object and its relational neighborhood. Apply the learned model templates to the target collection of linked objects. Neville and Jensen describe collective inference using repeated sampling of unknown variables from their local conditional distributions. Unlike an acyclic Bayesian network, a dependency network can contain cycles. Independently learned local conditionals may be inconsistent with one exact joint distribution, so inference has important qualifications. Sampling behavior, update choices and convergence checks matter; multiplying all local conditionals is not a general shortcut to a valid joint model.

Three practical examples

A hypothetical document-classification project estimates topics using both document content and links among documents, evaluating the predictions collectively.
A research team models attributes in a connected equipment dataset, asking whether information about related components improves prediction beyond each component's own measurements.
An analyst compares an RDN with an independent classifier to determine whether relational dependencies add useful evidence or merely spread errors among connected records.

Think of it this way

Think of several unfinished puzzles that share clues. A tentative answer in one puzzle can help another, which may then provide evidence back to the first. The analogy captures collective inference, but repeated agreement is not proof of correctness. If an early clue is misleading, connected guesses can reinforce the mistake rather than resolve it.

Worked example

Suppose two linked documents, A and B, have unknown binary topic labels. A toy local model assigns A probability 0.8 of being technical when B is currently technical, but 0.3 otherwise. Another local model assigns B probability 0.7 when A is technical and 0.2 otherwise. Starting with B nontechnical, a sampling update for A uses probability 0.3. If A is sampled as technical, the next update for B uses 0.7. Repeating this process illustrates conditional updates. One pass is not a final probability estimate, and these invented conditionals are not claimed to be mutually consistent or to describe real documents.

Illustrative case study

A fictional research-library team tests an RDN for classifying linked reports. It separates observed labels from labels to be predicted and records which links will actually be available in the intended application. The team compares content-only predictions with collective predictions and examines whether incorrect initial evidence spreads through tightly connected groups. It repeats inference runs, checks stability and evaluates on a deployment-relevant split. When relational information adds little beyond document text, the simpler model remains a serious alternative. The project treats network structure as potentially useful evidence, not as an automatic reason to trust connected predictions.

Common misconceptions and mistakes

An RDN is not a relational database schema or a database functional dependency diagram. It is not a Bayesian network with an arbitrary cycle added while retaining the same factorization rules. Locally plausible conditional models may fail to correspond to one coherent joint distribution. A converged sampling procedure does not guarantee that the model matches reality. Relationships can encode shared context, selection effects or recording practices rather than causal influence. Also avoid using true neighboring labels during testing if those labels would be unknown at deployment.

Business use

RDNs are relevant to specialist projects where dependencies among linked records may improve prediction. Start by defining which attributes and relationships are observed at decision time. Compare with a strong independent baseline and account for the cost of constructing the relational graph. Ask how collective inference is run, how stability is assessed and what happens when links or labels are missing. Evaluate errors at both object and connected-group levels. A small average gain may be insufficient if a single mistaken label can produce a cluster of operationally costly errors.

Check your understanding

In the toy example, why does B's update use 0.7 after A is sampled as technical?

Answer: B's local conditional model depends on A's current label.

Why is one update insufficient to report final uncertainty?

Answer: The state remains part of an iterative stochastic procedure.

What evaluation mistake occurs if unknown test-neighbor labels are supplied as facts?

Answer: The model receives information that may not be available in the actual task, inflating apparent performance.

Related terms

Learn next

Study ordinary dependency networks before moving to relational templates. Then compare independent classification with collective inference on a tiny graph. Examine conditional consistency, sampling diagnostics and deployment-aware label availability before interpreting larger network results.

Sources

Jennifer Neville and David Jensen, Relational Dependency Networks, Journal of Machine Learning Research: https://www.jmlr.org/papers/volume8/neville07a/neville07a.pdf Article record: https://www.jmlr.org/beta/papers/v8/neville07a.html The conditional probabilities and research-library case are original illustrations, not published experimental results. Independent review consulted Neville and Jensen's conditional-consistency discussion and section on RDN inference. These support iterative conditional resampling and the distinction between inconsistent local models and the sampler's stationary distribution. The original two-variable update was independently traced.

Previous: Relational database Next: Relational extraction