AI Glossary / AI and machine learning / Advanced
Analogical modeling
Definition
Analogical modeling, in its specialized sense, is Royal Skousen's exemplar-based approach to predicting behavior from stored examples and their contexts. It considers different generalizations of a target context and uses those that meet its formal homogeneity conditions to support a prediction.
What it means
People often predict a new case by recalling similar cases. Analogical modeling gives a particular formal treatment of that idea, especially in research on language behavior. It is more specific than the everyday expression “learning by analogy.”
An exemplar records a context and an observed outcome. Context features might describe surrounding sounds or other properties relevant to a linguistic choice. The method examines how examples behave when some of those features are retained and others are ignored.
The representation remains a modeling choice. Calling a method exemplar-based does not mean it avoids assumptions: someone still selects the features, examples and outcomes. A poor representation can hide the distinction that matters for the target behavior.
How it works
For a target case, the method constructs more general contexts, called supracontexts, by leaving selected features unspecified. With m binary choices of retaining or ignoring a feature, there can be 2^m such combinations before further handling.
Stored exemplars are associated with the contexts they match. The method distinguishes homogeneous from heterogeneous groupings according to its formal criteria involving outcomes and subcontexts. Accepted groupings contribute to the analogical prediction; rejected ones do not.
The original approach uses a weighting mechanism in which accepted contexts can contribute according to squared frequency. This is not the same as simply taking the nearest example or giving every matching row one equal vote. Detailed implementations need the exact homogeneity and weighting definitions. Computational growth can become substantial as the number of features increases.
Three practical examples
- A linguistics researcher investigates how recorded word forms support predictions about an unfamiliar form.
- A teaching exercise compares predictions based on exemplars with predictions from a manually written linguistic rule.
- An analyst studies how changing the selected context features changes which examples can support a target case.
Think of it this way
Think of sorting a collection of postcards by several attributes: destination, season and purpose. Different groupings may reveal useful regularities for choosing the next postcard.
The analogy does not reproduce the algorithm. Analogical modeling has exact rules for accepting groupings and weighting their evidence; simply choosing a group that looks persuasive would not implement the method.
Worked example
Counting possible context generalizations
Suppose a simplified target has three features: first symbol A, second symbol B and third symbol C. Each feature can be retained or ignored, giving 2 × 2 × 2 = 8 possible generalizations.
They include ABC, AB-, A-C, -BC, A--, -B-, --C and ---, where a dash means the feature is unspecified. An exemplar matching AB but differing on the last symbol can belong to AB- without matching ABC.
The eight contexts are only candidates. They are not eight equally valid votes. The algorithm still has to determine which contain relevant examples, which satisfy its homogeneity conditions and how accepted evidence is weighted. Increasing to ten features creates 1,024 combinations, illustrating why feature count affects computational work. These counts explain the context construction rather than a complete prediction.
Illustrative case study
A fictional language-technology research group studies a small collection of inflected word forms. It compares an analogical-modeling experiment with a simple rule-based baseline. At first, it treats spelling characters as the complete context.
The researchers then investigate whether a pronunciation-based representation changes the predictions. They keep the evaluation examples separate and report both successes and failures rather than selecting only convincing analogies.
The case illustrates the importance of representation in exemplar methods. It does not imply that one encoding or this particular algorithm is universally better for language tasks.
Common misconceptions and mistakes
- Analogical modeling means any use of examples. Here it names a particular formal approach.
- It is identical to nearest-neighbor classification. The context selection and weighting procedure are different.
- Exemplar-based methods eliminate feature engineering. The chosen representation still determines what similarities and differences the system can see.
Business use
The specialized method is mainly relevant to research and technical literacy. Its broader lesson for businesses is that reuse of past examples requires a defensible definition of context. Similar-looking cases may differ in the feature that actually determines the outcome.
If a supplier uses the term, ask whether it refers to Skousen's method or to analogy in a general sense. Then examine the implemented procedure, the examples available at prediction time and the evaluation on unfamiliar cases. The label alone does not establish a practical advantage.
People also ask
Who proposed this form of analogical modeling?
Royal Skousen developed the approach in work on language and exemplar-based prediction.
What is a supracontext?
It is a more general context produced by leaving some features of the target unspecified.
Does every supracontext contribute?
No. The method applies formal acceptance conditions before using a context's evidence.
Check your understanding
A target has three context features, each of which can be kept or ignored. Are the resulting eight combinations automatically eight equal votes?
Answer: No. They are candidate contexts. Their contents, homogeneity and weighting must still be evaluated.
Related terms
- Instance-based learning
- Nearest neighbor
- analogy
- exemplar-based learning
- feature representation
- linguistic modelling
Learn next
Read instance-based learning and feature representation. Describe a prediction problem using several possible context features, then consider which important distinctions each representation preserves or loses.
More in AI and machine learning
- Algorithmic transparency - Algorithmic transparency is the availability of understandable information about an algorithmic system's purpose, operation, data, limitations and use.
- Almeida - Pineda recurrent backpropagation - Almeida - Pineda recurrent backpropagation is a method for calculating learning gradients in recurrent neural networks that settle to a stable fixed point.
- Alternating decision tree - An alternating decision tree is a classification model that alternates between decision conditions and numerical prediction contributions.
- And - or tree - An and - or tree represents a problem using two kinds of branching.
- Anomaly - An anomaly is a data point or event that departs from what is normal in a way that deserves attention: the unusual transaction, the strange sensor reading...
- Anomaly detection - Anomaly detection is the task of finding data points, events or patterns that differ markedly from what is normal, such as a fraudulent payment, a failing...
Sources
Royal Skousen, Quantum Analogical Modeling with Homogeneous Pointers: https://arxiv.org/abs/1006.3308. Royal Skousen, Expanding Analogical Modeling into a General Theory of Language Prediction: https://citeseerx.ist.psu.edu/document?doi=e32c1bc5e1758a09b7afb8520a0b2741591344b3&repid=rep1&type=pdf. Author-written research references describe context homogeneity and weighting. The three-feature exercise is original and deliberately stops before a full prediction calculation.
Previous: Analog Filter Design Next: Analogical reasoning