What is Instance-based representation?

The simplest structure of learning is plain memorization, or rote learning. Because a group of training instances has been remembered, on encountering a new instance the memory is investigated for the training instance that most powerfully resembles the new one.

The only problem is how to clarify resembles. First, this is a completely different method of describing the “knowledge” extracted from a group of instances − It stores the instances themselves and works by associating new instances whose class is unknown to the current ones whose class is known. Rather than trying to make rules, work directly from the instances themselves. This is called instance-based learning.

In instance-based learning, all the actual work is completed when the time appears to define a new instance instead of when the training set is processed. The difference between this approach and the others that it can be seen is the time at which the “learning” takes place.

Instance-based learning is inactive, deferring the real work considering possible, whereas different methods are eager, generalizing as soon as the data has been seen. In instance-based classification, each new instance is distinguished from current ones using a distance metric, and the nearest existing instance is used to make the class to the new one. This is known as the nearest-neighbor classification method.

Sometimes more than one nearest neighbor is used, and the majority class of the nearest k neighbors (or the distance weighted average if the class is numeric) is created to the new instance. This is defined as the k-nearest-neighbor method.

When nominal attributes are current, it is essential to come up with a “distance” between multiple values of that attribute. Various attributes will be significant than others, and it is usually reflected in the distance metric by several types of attribute weighting. It is changing suitable attribute weights from the training group is an essential problem in instance-based learning.

An apparent limitation to instance-based representations is that they do not create explicit architecture that is learned. The instances connect with the distance metric to divide out boundaries into instanced areas that analyze one class from another, and this is a type of explicit description of knowledge.

For instance, given a single instance of each of two classes, the nearest-neighbor rule efficiently divides the instance area along the perpendicular bisector of the line connecting the instances. Given several instances of every class, the space is splitted by a set of lines that defines the perpendicular bisectors of selected lines linking an instance of one class to one of another class.