Must the data come from an SQL database?

No. The concept concerns relational structure, though multiple database tables are a common source.

Can aggregation still be useful?

Yes. Carefully chosen summaries can preserve useful information, but may discard details needed for some patterns.

Is this the same as graph learning?

There is overlap, but relational data mining is a broader historical family with several representations and algorithms.

AI Glossary / AI and machine learning / Advanced

Relational data mining

Definition

Relational data mining discovers patterns involving linked entities and relationships, often across multiple tables in a relational database. Instead of treating each observation only as an isolated row of attributes, it can use connections such as a customer placing an order or a machine receiving a component from a supplier.

What it means

Business information is commonly spread across tables because one entity can have many related records. A machine may have several inspections, repairs and installed parts. Relational mining looks for patterns involving those links and attributes together. Flattening everything into a single table is possible in many workflows, but careless joins can duplicate entities or obscure important relationships. The central question is not how many tables exist, but whether the model preserves the relational information needed to answer the intended question correctly.

How it works

Define the target entities, their attributes, the links between them and the prediction or discovery task. Search for patterns expressed through those relationships, or construct relational features for a suitable learner. Džeroski describes approaches including relational rules, trees and distance-based methods, with connections to inductive logic programming. Some methods operate directly on relational structure; others deliberately transform it into a feature table. Define aggregation and time boundaries carefully. Evaluation must account for linked observations, because related records can otherwise share information across training and test sets in misleading ways.

Three practical examples

A hypothetical maintenance team studies patterns connecting machines, component batches and service records to identify combinations associated with later inspection findings.
A supply-chain analyst examines whether delivery outcomes relate to supplier attributes and the characteristics of linked orders, keeping the order and supplier units distinct.
A research project represents molecules through atoms and bonds, investigating patterns that depend on connections rather than only a single summary value for each molecule.

Think of it this way

Think of reading a set of connected files rather than judging each sheet in isolation. A useful clue might appear only when a repair record is linked to the correct machine and component. The analogy also highlights a risk: connecting the wrong files can create a persuasive but false pattern. Relationships need reliable identities and clear meanings.

Worked example

Suppose the target is whether each of three machines requires follow-up. Machine A has four service records, B has one and C has one. Joining machine labels onto service rows creates six rows, with A appearing four times. If those rows are treated as independent machines, A receives four times the influence of B or C. A relational approach might instead ask whether a machine has any service involving a particular component, or count qualifying services before a defined date. The correct representation depends on the target question. This invented example illustrates why joins and aggregation need deliberate design rather than automatic flattening.

Illustrative case study

A fictional equipment distributor investigates recurring delivery problems across customers, orders and supplier records. Analysts first define the prediction point so only information available before dispatch can be used. They compare a customer-level summary model with a relational representation that preserves selected order and supplier connections. A validation review reveals that randomly splitting order rows lets the same tightly connected group appear on both sides, overstating the intended generalization. The team revises the evaluation to match future deployment and documents how missing links are handled. It accepts patterns as predictive evidence requiring review, not as proof that a supplier caused every associated problem.

Common misconceptions and mistakes

Relational data mining is not merely running any model on data exported from a relational database. The distinguishing feature is use of relationships in the patterns or representation. A database join does not automatically preserve the correct unit of analysis. Association across linked entities does not establish causation. Missing links may reflect incomplete recording rather than absence of a relationship. Information created after the prediction time can leak through related tables. Finally, highly expressive relational patterns can overfit, so additional flexibility requires appropriate evaluation and limits on the search.

Business use

The method is relevant when important signals lie in connected operational records. Begin with reliable entity identifiers and a clear definition of what is being predicted for whom and when. Compare direct relational modeling with a carefully engineered single-table baseline. Measure the cost of creating and refreshing links, not only model accuracy. Keep pattern explanations tied to actual relationship definitions so business users can assess them. Review whether evaluation represents new entities, new periods or new links, since these deployment settings can require different tests.

Check your understanding

Why does the six-row join not represent six independent machines?

Answer: Four rows belong to the same machine.

What should you decide before aggregating service records?

Answer: The target entity, relevant relationships and information available at prediction time.

Why might a random row split be misleading?

Answer: Linked records may transmit information between training and test portions in a way that will not be available in the intended future use.

Related terms

Inductive logic programming
Data leakage
Multi-relational data mining
relational learning
entity resolution
feature aggregation

Learn next

Map a small database's entities and one-to-many relationships. Then compare a joined table, entity-level aggregates and a relational rule for the same question. Study evaluation designs for linked and time-dependent records before interpreting a strong predictive score.

Sources

Sašo Džeroski, Multi-Relational Data Mining: An Introduction: https://research.cs.wisc.edu/EDAM/Dzeroski.pdf The machine-record calculation and equipment-distributor case are original illustrations. They do not report measured supplier performance or establish causal claims. Independent review consulted Džeroski's primary introduction and relational distance-based methods section. These support multi-table pattern discovery, relational rules/trees and distance-based approaches. The join example was independently traced with the machine as the target unit.

Previous: Relational data analysis Next: Relational Data Model