349.

[Cover] [Contents] [Index]

Page vii

Preface

We classify objects in order to make sense of our environment, by reducing a multiplicity of phenomena to a relatively small number of general classes. On a country walk, for example, you might point to cows, trees, tractors, or swans. What you are actually doing is to identify an observed object and to allocate it to a pre-existing class, or ‘give it a name’. Before setting out on the walk, you knew that swans existed, and you could specify their characteristics. When you saw a large white bird, possibly swimming in a canal or river, with an orange and black beak, you compared those characteristics to those of a swan and thus identified the bird, giving it the name or label of ‘swan’. We must be careful, therefore, to distinguish between the definition of the classes to which objects may belong and the identification or labelling of individual objects, and to avoid confusion between the two meanings of the word ‘classification’—i.e. the definition of categories of objects and the assignment or allocation of individual objects to these classes.

The example of the swan can also help to define other concepts. First, you must already have a model (or idealised representation) of the key features of a swan before you can recognise one. You learned, presumably in your childhood, the names of categories and subcategories of animals, plants and other objects. Now you use that knowledge to identify and name the things you see and hear. In the literature of classification, this approach is termed ‘supervised learning’, meaning that you have divided the phenomena of interest into a number of a priori groups, from each of which you have observed a number of examples, and have characterised them in terms of a number of discriminating features. The sample set is called ‘training data’, and this approach is known as ‘supervised classification’. In fact, it is supervised identification because it is assumed that the classification (the definition of the groups and their characteristics) has been defined before any previously unknown objects were identified.

An alternative approach, known as ‘unsupervised classification’ or ‘clustering’, is also widely used. In this approach it is assumed that you have little knowledge of the characteristics of the data set, that you wish to determine whether any natural groupings exist in those data and, if so, whether

[Cover] [Contents] [Index]