Browse Books

Reviewer: Friedrich Gebhardt

Books on data mining usually concentrate on results and on particular methods employed to achieve them. This one is different; it collects some methods that are essential for the field. This is a good idea. Unfortunately, the realization of the plan leaves something to be desired. After a short introduction to data mining in general, four chapters introduce basic techniques. The first one concerns rough sets. Here the reduction problem—finding minimal subsets of the variables that suffice for classification—takes much of the space. The chapter on fuzzy sets abounds with definitions. In the chapter on Bayesian methods, the authors emphasize classification and introduce probabilistic neural networks. The chapter on “Evolutionary Computing” gives a first introduction to genetic algorithms. These are the building blocks for the advanced methods in the next three chapters. Machine learning is used for discrimination between two or more groups by rules or decision trees; two pages discuss the application of this technique to knowledge discovery. The neural networks presented are the radial basis function network and Kohonen's self-organizing map network. In this book, clustering means finding representatives for the clusters that minimize some criterion; the authors use neural nets and fuzzy sets to achieve this goal. Data are mostly not as clean as is assumed by data mining methods; therefore, a chapter on preprocessing concludes the book. It centers on the use of principal component analysis and similar techniques to reduce the dimensions and create orthogonal variables. In all chapters, the theory is illustrated with many small examples and sometimes with larger examples. The exercises are adequate. Of course, a book on so many different subjects cannot replace a series of textbooks on each of them, and the selection of the material is a matter of taste. As the authors stress, the techniques described here presuppose that the variables are fairly independent; if they are not, orthogonal transformations should be applied in the preprocessing stage. Thus, methods whose purpose is to find such perhaps unknown dependencies (such as regression techniques) are not covered. While one could argue about the choice of subjects, the technical details are well below average. Misspellings and grammatical errors could be considered a nuisance, but numerous mistakes in the formulas (wrong indices, max instead of min, union instead of Cartesian product, inconsistent variable names, and so on) hinder comprehension. Some sections abound with definitions, while others dwell on undefined and poorly explained notions. The index could be expanded; even some terms that are used repeatedly are missing. Since the references are scattered over the chapters, an author index would be helpful.