Browse Books

Reviewer: Fernando Berzal

As the author highlights in his prologue, machine learning is the systematic study of algorithms and systems that improve their knowledge or performance with experience. Using a spam filter as an example, the reader is gently introduced to the "ingredients" of a machine learning system: tasks, models, and features. Machine learning tasks are stylized problems that can be solved with machine learning techniques, the most common of them being the classifier behind a spam filter, which distinguishes between spam (unwanted email) and ham (not spam). Two chapters are devoted to their description. The first one is focused on binary classification, as in the spam filter, and the related tasks of scoring and ranking. The second chapter extends the basic framework for classification problems with more than two classes and for the prediction of real-valued variables (regression). It also introduces other important machine learning tasks, including clustering, subgroup discovery, and association rules. Machine learning models are the output of machine learning algorithms that "make sense of data." Unsurprisingly, more than half of the chapters of Flach's textbook are dedicated to their study. The survey of available models starts with an overview of logical models, followed by two separate chapters on decision tree induction and rule induction, with discussions on ranking and probability estimation trees, descriptive rule learning for subgroup discovery, association rule mining, and learning first-order rules. After the 100 or so pages devoted to logical models, two chapters are devoted to geometric models. Geometric models encompass linear models (linear classifiers, linear regression, the perceptron, and support vector machines) and distance-based models (nearest-neighbor classifiers and the most common clustering techniques, for example k -means and hierarchical clustering). Short final sections nicely introduce the ideas behind kernel-based models that extend basic models and whose description is not always as readable as in this textbook. The final chapter on individual models is devoted to probabilistic models, of which na?ve Bayes and Gaussian mixture models are two prominent examples. After a detour through a separate chapter on features, a short 13-page chapter overviews model ensembles and completes a nice survey of existing machine learning models. Model ensembles combine individual models and often outperform them. In their chapter, readers will become acquainted with useful techniques such as bagging and boosting. Beyond the survey of machine learning models, the author devotes separate chapters to features and experiments. Features are the third ingredient of machine learning systems, and "the workhorses of machine learning," using the author's own words. Preprocessing them is often key in the performance of a machine learning system. Hence, many techniques have been proposed for transforming, selecting, and constructing features-discretization, normalization, calibration, and principal component analysis, to name just a few. The final chapter on experiments explores the importance of choosing the right performance measurement for the machine learning task at hand. Using a statistician's perspective, it takes a closer look at how to properly measure and interpret them. This textbook is clearly written and well organized. Starting from the basics, the author skillfully guides the reader through his learning process by providing useful facts and insight into the behavior of several machine learning techniques, as well as the high-level pseudocode of many key algorithms. Each chapter ends with a summary section that reinforces key points and provides just the right number of references for those interested in delving deeper. Summary sidebars are scattered throughout the book that review key mathematical concepts, making this book almost completely self-contained for novices. The book is also nicely typeset in color and includes plenty of cleverly designed figures that illustrate machine learning concepts as they are discussed in the text. There is only a minor caveat with respect to those figures: many of them contain so many details and their lines are so thin that they are hardly legible. It is surprising how much can be covered in around 400 pages. Did I mention the author's crystal clear explanations of computational learning theory (section 4.4, where concepts such as probably approximately correct (PAC) learning and the VC-dimension are introduced) and the minimum description length (section 9.5 on compression-based models), or his extensive use of coverage curves to visualize the behavior of machine learning models__?__ Obviously, breadth takes precedence over depth. The reader will often learn the intuition behind some models, but he will not be aware of all the technicalities that he should know in order to be able to implement the algorithms by himself. Of course, many might argue that well-tested software packages are available that provide the desired functionality. Hence, many readers would never try to implement their own machine learning algorithms. As users, however, they should always know the algorithms' theoretical foundations, key features, and, most of all, inherent limitations. This is precisely where this gentle introductory textbook excels. More reviews about this item: Amazon Online Computing Reviews Service