Document not found! Please try again

CART. Classification and regression trees- Leo ...

2 downloads 0 Views 112KB Size Report
Leo BREIMA'/, Jer(,me H. FRIEDMAN, Richard. A. OLSHEN ;,nd Cb.arles J. STONE. ('lassilicatioll and Recession Trees. The Wadsworth Statistics/Probabilit~.
14-;

Book Reviews

i i an exact method is not feasible for solving a particalar problem then an approximate method must be used; a general treatment of ~uch methods is Wen in Chapter 12. This is a topic rarely seen in text bc,oks. probably ix'cause approxi~aate reel hods b~, their very niture have a tendency to resist a satisfymg up.ified treatment, but the authors have su.zceeded and have provided a usefJI addition to the literature ia this area.

The boek i:~ very' readable but not always easy going. The exposition is good and clear perhaps resulting from the 'grass roots' experience of the authors who have produced many papers in the field. The references are extensive with many papers in French listed; this could be very useful to ,-.nglish speaking readers who might be less a,a are of such papers. The authors aim to provide a balance between *.heo~T and practice. I regard it as biased toward theory and I estimate the main readershit: will be among practically minded theorists. In summary, the a.~thors of the book and Steven Valda. who ~ranslated it. are to be complimented on a splendid and th,~roughly enjoyable text which de,,erves to ~gecome ;, ~tandard reference in ihe area. 7: B. BOFFE Y L mrersitl" of Lwerpool, Victoria Building Lit erpool. Untted Kingdom

Leo B R E I M A ' / , Jer(,me H. F R I E D M A N , Richard A. O L S H E N ;,nd Cb.arles J. S T O N E ('lassilicatioll and Recession Trees The Wadsworth Statistics/Probabilit~ Series. Wadsworth, Belmont, 1984, x + 358 pages in the classification problem measurements are made on some case or object and used to predict the cla:~s the case belongs to. The const,',action of a rule that assigns a class membership to every measurement vector is based on the measuremen t data on :\" cases, the learning sample, observed together with their actual classification. A well kv.own techaique for this type of problem is di:~criminaqt anal\'sis. l-his book considers the use of tree ,~ethods ii~ clz~,,,ification. Tree struc,a,ed classifier.; are con.strtected by repeated splits of subsets of the mca.surement space into two de:,cendant .,ubsets. A pro,-ess that can be depicted as a tree. Terminal subsets (those which are not :,;e,lit) are assigned to,

a class. Contrary to discriminant analysis no assumptions are made concerning underlying distributions and qualitative variables can be used. The essential questions in the construction of a tree classifier are - the choice of a splitting rule, - the decision to declare a subset terminal, - the assignment of a terminal subset to a class. These questions are studied in detail. The first five chapters, presenting the methodology of tree construction, are a wcll-writtep mixture of heuristics and theory. R e m a r k a b l e is the solution given f)r the second question, about the identification of terminal nodes. Instead of application of a s~cpping rule in the splitting process, here a two-step approach is proposed. First grow a tree that is much too large, and then prune it upward, according to a criterion taking account of both the c o m p l e d t y of the tree (i.e. the number of terminal ,redes) and the estimated misclassification rate. The results are more satisfactory than with the classical stopping rules. In Chapters 6 and 7 some examples are presented, and Chapter 8 is the only one devoted to the regression problem. The final four chapters present a theoretical framework, and can be read independently ef the earlier chapters. All in al~ ther,- i.- a good balance between theory and practical aspects and I think the book i, very ..seful for those '~,,ho are interested in the theory of tree classifiers, as well as :hose who have to cope v~,ith classification problems.

J. PKA A GMA N Emdhoven UnoersiJv of 7 echnologv 5600 MB Eindtwven. Netherlands" Lawrence A. G O R D O N and George E. '?INCHES Improving Cai;ital Budgeting: A Decis,.'on Support System Approach Library of Cc, ngress Cataloguing in Pubiication Data, Addison-Wesley, Reading, 1984, x + 1!6 pages Acc:-~rdiqg to Chapter 1 (p.5) this book i~, intended to p,ov:de a link between decision support s3/:,tems a!~.d ,::~;:.ital budgeting. It dividea the deci.~i~m-making pr'~ee.,:s in four ;~hase.s: --- identification: of opportunities.

Suggest Documents