exemplarbased

Exemplar-based, also known as exemplar-based learning or instance-based learning, refers to a family of approaches in machine learning and pattern recognition that make predictions by comparing a new example to a stored collection of labeled instances, or exemplars. These methods are typically non-parametric and lazy: they do not build a global model during training, but defer generalization until a query is made, using the exemplars as the basis for decision.

Prediction is usually accomplished by measuring similarity between the query and exemplars using a distance or similarity metric. The most common example is k-nearest neighbors (k-NN), where the k closest exemplars determine the class by majority vote or a weighted vote, or a predicted value by weighted averaging. Case-based reasoning (CBR) follows a cycle of retrieving a similar past case, reusing its solution, revising as needed, and retaining the new case for future problems.

Advantages of exemplar-based methods include simplicity, flexibility, and the ability to model complex, non-linear decision boundaries without assuming a particular parametric form. They adapt naturally to multimodal distributions and can work with a wide range of data representations, provided an effective distance metric is available.

Disadvantages include computational and storage demands at query time, sensitivity to irrelevant or noisy features, and the curse of dimensionality in high-dimensional spaces. Performance depends heavily on choosing appropriate features and metrics, as well as effective instance selection and potential metric learning.

Exemplar-based approaches are used in image and text classification, information retrieval, and medical decision support, and are often contrasted with prototype-based or model-based methods that rely on summarized representations or learned parameters.