Classical methods of discrimination

Publication date 18 November 2001

Bayesian discrimination rule discriminant analysis kernel density estimation maximum likelihood rule symbolic objects

Bayesian inference (62F15) Density estimation (62G07) Classification and discrimination; cluster analysis (statistical aspects) (62H30)

Review text

The identification of the class from which an observation \(x\in\mathcal B\) originated requires the definition of a mapping \(d\), called a decision rule: \(d :\;{\mathcal B}\to\{1,\dots,g\} (x\to d(x))\). This rule depends on the training data \(x_{11},\dots,x_{1n_1},\dots,x_{g1},\dots,x_{gn_g}\) and is typically constructed in view of the minimization of an overall error rate. Each rule \(d\) determines a partition \((P_1,\dots,P_g)\) of the space \(\mathcal B\) of possible data values: \(P_k=\{x |d(x)=k\},\;1\leq k\leq g.\) The rule \(d\) is often defined with the help of discriminant functions \(h_k(x),\;k=1,\dots,g\). The function \(h_k(x)\) is associated with the \(k\)-th population \(\Pi_k\) and gives an idea of the ``similarity which exists between \(x\) and the individuals from the \(k\)-th training sample \(G_k=\{x_{k1}, \dots,x_{kn_k}\}\). For classical data with \({\mathcal B}={\mathcal X}=\mathbb{R}^p\) the discriminant functions can be determined by probabilistic approaches. The classical Bayesian discrimination rule gives a solution to the discriminant analysis problem for classical data. As for most classification methods, this discrimination rule needs to be adapted to the symbolic objects problem in which we can have information of different types: quantitative, qualitative, probabilistic, interval-type. To perform such an adaptation, some ``density measure is needed on the symbolic objects observation space \(\mathcal X\). A generalization of the kernels method provides this ``density estimation where ``density measures the concentration of data points in the neighbourhood of a fixed data point and has here no probabilistic meaning. The author gives a short review of the classical methods and results (maximum likelihood rule, Bayesian rule, density estimation methods) that are generalized in the following articles of this book in order to develop discriminant analysis for symbolic objects. For the entire collection see [Zbl 1039.62501].

Recommendations

Cited in

(8)

This page was built for publication: Classical methods of discrimination

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2756644)