Can a dependency network contain cycles?

Yes. That is a defining contrast with an ordinary Bayesian network.

Is it a project dependency chart?

No. Here it is a probabilistic graphical model of variables.

Can it support collaborative filtering?

Yes. Preference prediction was one of the applications in the original research.

AI Glossary / AI and machine learning / Advanced

Dependency network (graphical model)

Definition

A dependency network is a probabilistic graphical model that describes each variable through a conditional distribution given selected other variables. Unlike a Bayesian network, its directed graph may contain cycles, making it useful for representing predictive dependencies without assuming a causal ordering.

What it means

A business dataset can contain several attributes that help predict one another. Product preferences, for example, may be informative in both directions: liking one item can help predict liking another, and vice versa.

A dependency network provides a collection of local predictive models for those relationships. Its arrows indicate which variables enter a conditional model, not necessarily which variables cause others.

The distinction from a Bayesian network matters. A Bayesian network's acyclic structure supports a particular joint-probability factorization. Independently learned conditionals in a dependency network need not be mutually consistent with one joint distribution. That issue affects how inference should be understood.

How it works

For each variable, a learning procedure estimates a conditional model using other selected variables. The resulting graph can contain directed cycles because each local prediction problem is learned without requiring a global acyclic order.

Inference can use a Gibbs-sampling-style procedure that repeatedly updates variables from their conditional models. Conditions for convergence and consistency require attention; the collection should not be treated as an arbitrary set of probabilities that automatically fits together perfectly.

Evaluation depends on the intended task, such as predicting preferences or visualizing dependencies. Local predictive accuracy and the behavior of the full inference process should be distinguished.

Three practical examples

A recommendation research system models one product preference using a user's other recorded preferences.
An analyst explores predictive relationships among several categorical survey variables.
A probabilistic-modeling project compares locally learned dependency models with an acyclic Bayesian-network approach.

Think of it this way

Think of several specialists, each predicting one part of a report from the other parts. Their predictions can inform one another in a loop.

The analogy has limits because individually plausible forecasts may not form one coherent joint account. The mathematical model needs an explicit inference procedure and checks of its assumptions.

Worked example

Predictive arrows do not establish causal direction

Imagine a fictional preference dataset with variables for liking hiking guides, outdoor equipment and travel books. A local model for outdoor equipment uses hiking-guide preference. Another local model for hiking guides uses outdoor-equipment preference.

The resulting graph can contain arrows in both directions. This represents mutual predictive usefulness. It does not prove that purchasing a hiking guide causes someone to like equipment or that buying equipment causes an interest in guides.

Both may reflect an unrecorded interest in outdoor activities. A recommender may still benefit from the association, provided it is evaluated on relevant held-out preferences.

If the business instead asks whether promoting a guide will increase equipment purchases, it has asked a causal intervention question. The dependency graph alone does not answer it. The example explains the meaning of the arrows without claiming a specific trained model or published result.

Illustrative case study

A fictional media service studies a dependency-network recommender. Its local models predict each content preference from other known preferences and appear strong in retrospective evaluation.

The team discovers that some test preferences were recorded after the recommendation time. It redesigns evaluation to reproduce the actual information available to the service and separately checks the iterative inference process.

The resulting report describes predictive dependencies and explicitly avoids causal claims about what content creates an interest. The case illustrates that a flexible graphical representation still requires careful time boundaries and a clear definition of the prediction task.

Common misconceptions and mistakes

Treating arrows as causes: they describe inputs to conditional models in this context.
Assuming cycles make the model a Bayesian network: ordinary Bayesian networks are acyclic.
Assuming all local conditionals are automatically jointly consistent: independently learned models can disagree.
Multiplying conditionals as though the Bayesian-network factorization applied unchanged: the representation requires its own inference treatment.

Business use

Dependency networks are mainly relevant to specialized probabilistic modeling, recommendation research and exploratory dependency visualization. They can make local predictive relationships easier to learn and inspect.

Ask how each conditional model was estimated, what inference procedure is used and whether consistency assumptions matter for the claimed result. Evaluate recommendations using information available at the actual decision time. If a diagram is shown to business users, label it as predictive so it is not mistaken for a causal process map.

Check your understanding

A dependency network contains arrows both from A to B and from B to A. Does it prove a two-way causal feedback process?

Answer: No. The arrows can simply show that each variable helps predict the other. Causal conclusions require additional assumptions and evidence.

Related terms

Learn next

Read conditional probability and Bayesian network. For one graphical model, identify whether arrows mean predictive inputs, causal effects or something else before interpreting the diagram.

Sources

Heckerman and colleagues, Dependency Networks for Inference, Collaborative Filtering, and Data Visualization: https://jmlr.org/papers/volume1/heckerman00a/heckerman00a.pdf. The abstract, introduction and section 3.1 were checked for cyclic predictive graphs, potentially inconsistent local conditionals and ordered pseudo-Gibbs inference assumptions. Examples are original.

Previous: Dependency grammar Next: Dependency tree