AI Glossary / Graphs and symbolic AI / Intermediate

Generalization hierarchy

Definition

A generalization hierarchy organizes concepts or descriptions from more specific to more general. In symbolic machine learning, it can help a learner replace detailed alternatives with broader predicates while keeping their meaning explicit.

What it means

A specific description applies to a narrower set of cases. A more general description applies to every case covered by that specific description and potentially to additional cases. For example, in a defined equipment catalog, laser printer may be a kind of printer, and printer a kind of output device.

The hierarchy supplies background knowledge for moving between these levels. It can reduce the need to express a rule as a long list of separate alternatives. However, a broader label may introduce cases that were absent from the original list, so shortening a rule and preserving its coverage are different achievements.

This entry uses the concept-learning sense. The same phrase also appears in data aggregation and privacy work, where the operational purpose and requirements can differ.

How it works

Define the concepts and the more-specific-than relationships between them. Check that each proposed relationship has the intended inclusion meaning. A relation such as manufactured-by is not a generalization link: a device is not a kind of its manufacturer.

A learner can then search upward for broader descriptions that cover positive examples, or downward for more specific descriptions that exclude negative examples. Training examples and existing candidate rules constrain which moves are acceptable.

Some systems organize not only attribute values but also relations and complete concept descriptions. The structure need not be a simple tree; a concept may have more than one valid generalization. Crucially, being higher in the hierarchy does not make a concept a better prediction rule. The rule must still be checked against its task and evidence.

Three practical examples

• An equipment classifier considers whether several printer models can be represented by a shared printer category.
• A symbolic troubleshooting system learns rules at a component-family level while retaining exceptions for individual components.
• A document-tagging prototype uses a topic hierarchy to compare a narrow rule with a broader rule that covers additional document types.

Think of it this way

Think of labeled shelves inside larger labeled sections of a store. A shelf for fountain pens belongs within writing instruments, so the broader section includes all of that shelf and more.

The analogy illustrates inclusion. Store layouts sometimes group products for convenience rather than logical membership; only genuinely inclusive relationships support the reasoning required by a generalization hierarchy.

Worked example

Checking what a broader label adds

Consider an invented catalog with four mutually exclusive device types: laser printer, inkjet printer, scanner and monitor. Its hierarchy defines Printer as laser printer or inkjet printer. OfficeDevice contains all four types.

A rule currently flags a device when it is a laser printer or an inkjet printer. Replacing those two alternatives with Printer preserves the rule's coverage under the stated definitions. Replacing them with OfficeDevice does not: scanners and monitors would now also be flagged.

Suppose a test set contains 3 laser printers, 2 inkjet printers, 4 scanners and 1 monitor. The original rule and the Printer rule each flag 5 cases. The OfficeDevice rule flags all 10. If only printers should be flagged, that broader rule introduces 5 false positives.

Now add a new subtype, thermal printer, under Printer. The two-item list still excludes it, while the Printer rule includes it. Whether that change is desirable depends on the target concept. The hierarchy makes the difference visible; it cannot decide the business meaning by itself.

Illustrative case study

The fictional Cedar Catalog team develops a rule that identifies equipment eligible for a printer maintenance workflow. Its first examples mention only laser and inkjet devices, and the team considers using a broader catalog label.

A reviewer checks the hierarchy and discovers that the proposed label also includes scanners. The team chooses Printer instead, then asks the service owner how thermal printers should be treated. It records that decision in the eligibility rule rather than assuming every catalog ancestor is suitable.

Cedar retains examples from the excluded neighboring categories as regression checks. The case shows how domain structure can help rule learning while still requiring explicit target definitions. It does not report a measured improvement in classification accuracy.

Common misconceptions and mistakes

• A generalization hierarchy is merely a list of related words. Its relevant links express a more-general or more-specific relationship.
• A shorter rule always has the same meaning. A broader concept may include additional cases.
• The hierarchy proves that a learned rule is correct. It supplies background structure; examples and task definitions remain necessary.
• Every hierarchy must be a tree. Multiple generalizations can require a more general partial-order structure.

Business use

The practical value is a more manageable vocabulary for rules, explanations and knowledge maintenance. A business can discuss eligibility at a meaningful category level instead of maintaining a growing collection of nearly identical conditions.

The main review question is coverage: which cases become included or excluded when a rule moves between levels? Keep category definitions and rule purposes separate. Updating a taxonomy can silently change rules that depend on its ancestors, so evaluate new subtypes and revised relationships against representative positive and negative cases. A clear hierarchy supports review; an ambiguous one can spread errors across many rules.

People also ask

Is this the same as a decision tree?

No. A decision tree branches on tests to make predictions. A generalization hierarchy organizes descriptions by relative generality.

Can a concept have several parents?

Yes, where each parent expresses a valid broader concept and the representation permits multiple inheritance.

Does moving upward always improve generalization to new data?

No. It increases coverage, but that additional coverage may include incorrect cases.

Check your understanding

Why does replacing the original two printer types with OfficeDevice change the worked rule?

Answer: OfficeDevice also includes scanners and monitors. The replacement expands the accepted set beyond the original alternatives, creating five false positives in the stated test set.

Related terms

• Hypothesis space
• Version space
• Generalization
• Specialization
• Taxonomy
• Predicate (logic)
• Inductive bias
• Concept learning
• predicate

Learn next

Draw a hierarchy for five familiar product types. Write a rule using two specific types, then replace them with a shared ancestor. List every newly included type and decide whether the replacement preserves the intended rule.

More in Graphs and symbolic AI

• Game theory - Game theory studies situations in which the outcome for each decision-maker depends on choices made by others as well as their own.
• Game tree - A game tree represents the possible sequences of decisions in a game.
• Gaussian process - A Gaussian process is a collection of random variables for which every finite collection has a joint Gaussian distribution.
• Generalization loss - Generalization loss is a model's expected loss on examples drawn from the target data distribution.
• Generalized modus ponens - Generalized modus ponens is a first-order logic inference rule that derives a conclusion when known facts match all premises of an implication under a...
• Genetic algorithms - Genetic algorithms are population-based search methods inspired by biological evolution.

Sources

Russell's author-provided machine-learning chapter discusses using broader predicates to avoid unwieldy disjunctions in concept learning: https://aima.eecs.berkeley.edu/~russell/papers/hpc-machine-learning.pdf

Kerber's original paper, Using a Generalization Hierarchy to Learn from Examples, describes the OTIS system and its hierarchy of descriptions: https://api.pageplace.de/preview/DT0400.9781483297699_A23889470/preview-9781483297699_A23889470.pdf

Relevant returned source text was reviewed. The equipment hierarchy, counts and Cedar Catalog case are original teaching examples.

Previous: Generalization gap Next: Generalization loss