This book presents the pattern recognition problem on the basis of learning methods. It focuses on the use of “dissimilarity” between objects. The book’s main thesis is that objects to recognize cannot be merely represented as parameter vectors in finite-dimensional spaces with Euclidean distance, but that the choice of the distance (or pseudo-distance) is also a crucial part of the representation. So, the learning base must include information about what are the similar and dissimilar objects and a quantification of these facts. In consequence, the choice of a (possibly Euclidean) distance on the objects representation set is related with a (not necessarily linear) embedding in a big (linear) space. The book is divided into 2 parts and 12 chapters. The first part is theory, the second is practice.
(1) After a general introduction, chapters 2 and 3 are devoted to quasi-topological spaces and characterization of dissimilarity measures.
(2) Chapter 4 presents learning approaches (statistical and inductive), with emphasis on the role of dissimilarity measures.
(3) Chapter 5 shows the variety and universality of dissimilarity information used in the literature.
(4) In chapters 6 and 7, the second part starts by presenting data exploration through visual tools, reduction by clustering, or dimensionality analysis.
(5) Chapters 8 to 10 describe classifiers, beginning with the One-Class problem, ending by combination methods.
(6) Chapter 11 surveys representations and gives recommendations for their use.
(7) Not surprisingly, the book ends by conclusions and open matters.