Must the rule be symbolic?

No; numerical models can learn inductively.

Does more data guarantee correctness?

No; relevance, quality, assumptions, and coverage matter.

Can several rules fit the same data?

Yes, as the example demonstrates.

AI Glossary / Graphs and symbolic AI / Advanced

Inductive learning

Definition

Inductive learning is learning a general rule or predictive function from specific examples. In the input-output setting used here, the learner observes inputs paired with known outputs and constructs a hypothesis for other inputs. The resulting rule may be approximate, and agreement with the observed examples does not guarantee correctness on new cases.

What it means

A learner might see descriptions of past support requests paired with their assigned departments. It then attempts to predict the appropriate department for another request. The important step is extending beyond the particular records already observed.
This differs from deducing a conclusion that follows necessarily from stated premises. Inductive predictions depend on assumptions about which patterns are likely to continue. Those assumptions can be useful without being logically guaranteed by the data.
This entry emphasizes the supervised input-output sense in the supplied glossary rather than every broader use of induction.

How it works

First specify the inputs, the outputs to be predicted, and the examples available for learning. Decide which candidate functions the learning procedure can represent; this collection is its hypothesis space.
The procedure searches or adjusts candidates using the training examples. The fitting criterion and the model's restrictions influence which candidate is selected.
Different hypotheses can agree on observed examples and disagree elsewhere. A learner therefore needs an inductive bias: restrictions or preferences that guide its choices beyond the data. Preferring a simpler rule is one possible choice, not a guarantee that the simpler rule describes the world.
Evaluation asks how well predictions work beyond the fitting examples. Poor measurements, inconsistent labels, omitted information, and changes in the operating environment can all undermine useful generalization.

Three practical examples

A support team learns request categories from previously labeled messages. A useful model should handle new wording, rather than merely recognize copied training messages.
A packaging team learns a handling-time estimate from recorded item features and measured completion times. A prediction for a new package extends a pattern beyond the observed records.
A maintenance team learns to classify inspection photographs from labeled examples. It checks whether performance changes across camera settings because an incidental background pattern might explain apparent training success.

Think of it this way

Think of learning which parcels need extra packing by observing past decisions. You form a rule that helps with the next parcel. Several rules may explain the examples you have seen, so a new combination of features can expose what the earlier examples left unresolved.

Worked example

Consider an invented binary classification task. Each parcel has two binary features: F means fragile and H means heavy. Output 1 means extra packing, and output 0 means ordinary packing.
The initial training set contains three distinct feature combinations:
F = 0, H = 0 gives output 0.
F = 1, H = 0 gives output 1.
F = 1, H = 1 gives output 1.
Compare two candidate rules. Rule A predicts extra packing exactly when F = 1. Rule B predicts extra packing when either F = 1 or H = 1.
On the three training cases, both rules predict 0, 1, 1. Both therefore have zero training mistakes and training accuracy of 3/3, or 100%. This score cannot distinguish them.
Now consider the missing combination F = 0, H = 1. Rule A predicts 0, whereas Rule B predicts 1. The existing training labels do not determine which answer is correct. A preference for using fewer features might select A, but that preference adds something beyond the examples themselves.
Suppose a reliable label is obtained for the missing combination and its output is 1. Under the assumption that labels describe a deterministic function of these two features, A is now inconsistent with the expanded data, while B remains consistent. If the learner is restricted to these two candidate rules, B is the sole surviving candidate.
For a separate arithmetic illustration, suppose all four combinations are equally likely in a fixed population and B is its true labeling rule. A is correct on three combinations and wrong on one, so its population accuracy is 75%; B's is 100%. This population calculation follows from the stated hypothetical distribution and rule. It is not a conclusion justified by the original three examples alone.
If instead A were the true population rule, the same original observations would still fit both candidates, but their population accuracies would reverse. A would score 100% and B 75%. The training evidence leaves these two possible worlds unresolved.
Finally, the claim of complete identification depends on the declared domain. Four labels cover all combinations of two binary features for a deterministic target. They do not establish a universal packing rule if actual decisions also depend on an unrecorded feature, such as package condition, or if the target changes over time.

Illustrative case study

Fictional Willow Packing develops a model from historical handling decisions. Fragile parcels dominate the examples requiring extra packing, and the initial model appears accurate.
A reviewer asks about heavy parcels that are not fragile. This combination is absent from the development sample. The team gathers labels for that gap and records the operational rule used by the reviewers.
Willow treats the new evidence as a reason to reassess the model.

Common misconceptions and mistakes

“Perfect training accuracy proves the rule.” Unseen cases may distinguish equally fitting hypotheses.
“Induction is deduction.” Inductive conclusions extend beyond what the examples entail.
“Bias always means unfairness.” Inductive bias describes learning assumptions and preferences.
“More copies resolve ambiguity.” Repeating the same cases may leave the missing combination unresolved.
“A learned relationship proves causation.” Predictive success alone does not establish a causal explanation.

Business use

Define the decision the prediction supports and examine gaps in the examples. Record label meaning, feature availability, and changes in operating conditions. Ask where plausible rules disagree, then seek relevant evidence. Judge the cost of mistakes as well as average accuracy: two models with similar scores may fail on different operational cases.

Check your understanding

Why do A and B tie initially?

Answer: They agree on all three observed combinations.

Which case separates them?

Answer: A heavy parcel that is not fragile.

Where does the 75% figure come from?

Answer: The hypothetical uniform four-case population.

Related terms

Learn next

Change the probability of the missing feature combination to 10%. Under the same deterministic target B, calculate A's population accuracy and explain why training accuracy remains unchanged.

Sources

AIMA glossary, inductive learning: https://github.com/aimacode/aima-glossary/blob/master/glossary.md Tom Mitchell, concept learning and inductive bias: https://www.cs.cmu.edu/afs/cs.cmu.edu/project/theo-20/www/mlbook/ch2.pdf University of Cambridge, supervised learning and generalization: https://www.cl.cam.ac.uk/teaching/0809/ArtIntI/notes4.pdf

Previous: Inductive bias Next: Inductive logic