Are these networks non-probabilistic because they use signs?

No. The signs stand for constraints on probabilities.

Does a plus sign mean the probability rises by a fixed amount?

No. It specifies a direction under stated conditions, not a magnitude.

Can a qualitative network return “unknown”?

Yes. That can be the correct result when competing influences or missing constraints prevent a justified directional conclusion.

AI Glossary / Graphs and symbolic AI / Intermediate

Qualitative probabilistic networks

Definition

Qualitative probabilistic networks represent probabilistic relationships using qualitative constraints, such as whether one variable has a positive, negative, zero, or unknown influence on another. They abstract away many exact probability values while retaining some information about how probabilities can change. They are useful when directions of influence are better supported than precise numerical estimates.

What it means

A conventional Bayesian network specifies conditional probability distributions. A qualitative version can instead record orderings among relevant conditional probabilities. For binary variables, a positive local influence can mean that the probability of the child being true is no smaller when the parent is true than when it is false, holding the other parent values fixed.
This is a statement about a probability comparison, not a promise that the child will become true. Positive influence also need not be large or strictly positive in every context. Different qualitative assertions may be compatible with many numerical networks, so the model usually leaves exact probabilities unresolved.

How it works

Define the variables, their ordering or true/false interpretation, and the network's probabilistic structure. Specify qualitative influence constraints and any context restrictions. Apply inference rules justified by those constraints and the relevant conditional independencies.
Along an appropriate simple influence chain, signs combine like multiplication: two positive influences give a positive overall influence, while a positive and a negative influence give a negative one. Competing parallel influences can create ambiguity; a positive and a negative route do not generally cancel to zero. Some formalisms include additional synergy information. The network structure and evidence matter, so sign manipulation must follow the chosen probabilistic semantics rather than an informal diagram alone.

Three practical examples

A process-modeling workshop can record which conditions tend to raise or lower the probability of a delay when precise rates are unavailable.
A research prototype can identify which uncertain relationships prevent a clear qualitative conclusion and guide further data collection.
A teaching tool can compare a numerical Bayesian network with a qualitative abstraction to show what information is preserved and what is lost.

Think of it this way

Think of a map of influences that says “pushes upward,” “pushes downward,” or “no established direction,” without labeling every push's strength. The analogy is limited because the arrows concern conditional probabilities. They must satisfy a formal model, and opposing arrows do not tell you which effect dominates without more information.

Worked example

Consider an invented binary chain A → B → C with no other paths. Let A mean a document is complex, B mean it requires extra review, and C mean it enters a later processing batch. Assume C is conditionally independent of A given B.
For a numerical check of one model compatible with positive signs, set P(B true given A true) = 0.8 and P(B true given A false) = 0.3. Set P(C true given B true) = 0.9 and P(C true given B false) = 0.2. Both local influences are positive.
Marginalizing B gives P(C true given A true) = 0.9 × 0.8 + 0.2 × 0.2 = 0.76. For A false, the result is 0.9 × 0.3 + 0.2 × 0.7 = 0.41. The increase is 0.35, confirming a positive overall direction in this particular numerical realization.
A qualitative network retaining only the two positive signs can support the direction under the stated chain assumptions, but it cannot recover the exact values 0.76, 0.41, or 0.35. Those require the numerical probabilities supplied for this illustration. Adding a separate opposing influence could make the net direction unresolved.

Illustrative case study

Fictional case: Juniper Operations sketches a network about document turnaround. Staff agree that additional review makes a later batch more likely, but disagree about its magnitude. They also identify an expedited route that can offset the delay for some documents.
The team records the opposing influences instead of inventing a single precise percentage. It identifies which conditional rates would help resolve the uncertainty. This fictional case illustrates how qualitative modeling can expose a trade-off and a data need. It does not report a real causal discovery, measured improvement, or validated forecasting result.

Common misconceptions and mistakes

Do not confuse a positive influence with certainty, a large effect, or a universal causal claim. A probabilistic dependency graph needs additional causal interpretation before intervention conclusions are justified. A relationship may change sign across contexts, which should be represented rather than hidden behind one unconditional plus sign.
Do not assume that two opposite influences sum to zero. With unknown magnitudes, the result is generally ambiguous. Nor does qualitative information eliminate the need to model dependencies correctly. Conditioning on evidence can change which paths transmit information. A numerical answer invented to resolve an ambiguous sign is additional modeling input, not a conclusion supplied by the qualitative network.

Business use

The practical value is a disciplined way to represent partial probabilistic knowledge. Teams can document justified directions, separate them from unsupported magnitudes, and identify the relationships that matter to a decision.
For a business pilot, give every node a clear meaning and ask whether a sign holds across the relevant contexts. Preserve disagreements and exceptions explicitly. Use a numerical model when the decision needs expected counts, detailed trade-offs, or calibrated probabilities that the qualitative description cannot provide. The qualitative model can guide that work, but it should not be presented as delivering precision its inputs do not contain.

Check your understanding

Can the two positive signs in the worked chain alone establish that C's probability increases by exactly 0.35?

Answer: No. They support a qualitative direction under the chain assumptions. The numerical increase 0.35 comes from the specific conditional probabilities used in the check; other compatible tables can produce different increases.

Related terms

Bayesian network
Conditional independence
Probabilistic inference
Sensitivity analysis
qualitative influence
sign propagation
synergy
influence diagram

Learn next

Learn conditional independence and qualitative trade-off resolution next. Change P(C true given B true) to 0.6 and the value given B false to 0.4. Recompute the two C probabilities and observe how the positive direction remains while its numerical strength changes.

Conditional independence

Sources

Michael Wellman, Qualitative Probabilistic Networks for Planning Under Uncertainty, original paper: https://arxiv.org/pdf/1304.3115 Michael Wellman, Fundamental concepts of qualitative probabilistic networks, original paper abstract: https://www.sciencedirect.com/science/article/pii/000437029090026V

Previous: Qualitative physics Next: Quality of Data