Classification methods, also known as classification procedures, are methods and criteria for classifying objects or situations into classes, i.e. for classification. Such a procedure is also known as a classifier. Many methods can be implemented as algorithms; this is also referred to as machine or automatic classification. Classification methods are always application-oriented, so there are many different methods.
In a narrow sense, in contrast to the classic classification methods, there are classification methods, which serve to classify objects into already existing classes. Colloquially, however, no distinction is made between them. Classification methods play a role in pattern recognition, artificial intelligence, documentation science and information retrieval, among others. To assess a classifier, various parameters can be determined. Since a strictly hierarchical classification of classification methods is hardly possible, they can best be classified on the basis of various characteristics:
- Manual and automatic procedures
- Numerical and non-numerical methods
- Statistical and distribution-free methods
- Supervised and non-supervised procedures
- Fixed-dimensioned and learning processes
- Parametric and nonparametric methods
In the case of automatic methods, the classification takes place by means of an automatic process by software. The process of machine classification can be described as a formal method of making decisions in new situations based on learned structures. Machine classification is a subfield of machine learning.
More precisely, this is the generation of an algorithm (the learning algorithm) that calculates structures when applied to known and already classified cases (the database). These newly learned structures enable another algorithm (the evaluating algorithm) to assign a new and previously unknown case to one of the known target classes based on the observed attributes and their characteristics.
Statistical methods are based on density calculations and probabilities, while distribution-free methods use clear dividers to separate the classes. The boundaries between the individual classes in the characteristic space can be specified by a discriminant function.
Examples of statistical methods include the Bayesian classifier, the fuzzy pattern classifier, and the nuclear density estimator. The calculation of separation surfaces is possible by so-called support vector machines.
The creation of structures from existing data is also referred to as pattern recognition, discrimination or supervised learning. Class divisions are specified, which can also be done by random sampling. In contrast, there is unsupervised learning, in which the classes of the data are not predetermined, but must also be learned. However, reinforcement learning can add information about whether a class division was correct or wrong. An example of unsupervised methods is cluster analysis.
Parametric methods are based on parametric probability densities, while nonparametric methods (e.g. nearest neighbor classification) are based on local density calculations.