Browse Books

Reviewer: Mirjana K. Ivanovic

It seems that the answer to the old question of whether mixing labeled and unlabeled data can produce better models in machine learning is converging to a definite “yes.” However, the assumptions, conditions, and techniques that make it possible are still areas of active research within the field of semisupervised learning (SSL). The field has been driven forward in recent years by the increasing availability of data, and also by the proportionally increasing costs of its preprocessing and labeling. As the first book-length treatment of SSL, this volume is a milestone. The editors admit that, as a field, SSL is not yet mature enough to produce a well-researched monograph that would serve both researchers and practitioners as a definitive reference. Instead, they collect practically all seminal and relevant developments, contributed by leading researchers in the field. Moreover, the editors take great care to ensure that the 25 chapters, written by 52 authors, complement each other by mutual referencing, use of consistent notation, and an overall organization that tells a story. The introductory chapter of the book presents a concise and clear discussion of what semisupervised learning is, its history, and the assumptions and settings that drive SSL, namely, the smoothness assumption, the cluster assumption (equivalent to low density separation), the manifold assumption, and transduction. The remaining chapters are organized into parts that roughly correspond to the underlying assumptions. Part 2 deals with methods that directly implement low density separation, Part 3 addresses graph-based methods that can be viewed as implicitly complying with the manifold assumption, and Part 4 covers algorithms that perform two-step learning via a change of representation that preserves the smoothness assumption. Part 1, dealing with generative models, has been put to the front since generative approaches naturally extend into the semisupervised setting, providing a baseline for discussions throughout the book. Chapter 2, which begins Part 1, provides a further taxonomy of SSL techniques by their probabilistic underpinnings, distinguishing between the generative paradigm, the diagnostic paradigm, and input-dependent regularization. The two remaining parts of the book, 5 and 6, turn to applications and perspectives of semisupervised learning, respectively. An outstanding feature of Part 5, and the whole book, is the introduction of a diverse and comprehensive benchmark (chapter 21), which should make the task of comparing experimental results easier in the future. Part 6 introduces an approach to SSL that seems to be independent of the previously discussed assumptions (chapter 23), and since SSL is a highly practical field, presents how SSL methods may be analyzed within important existing theoretical frameworks. Another outstanding feature of the book is the concluding chapter, which is written in the form of a hypothetical debate between three researchers about the differences between semisupervised and transductive learning, bringing together many of the important points discussed in previous chapters. The book is primarily targeted at researchers and advanced practitioners, and graduate-level knowledge of probability, statistics, linear algebra, and calculus is required to fully comprehend its content. However, the contributions are very well written and, together with the introductory chapters, provide discussion that may be useful to practitioners who do not wish to delve into the gritty details. The book should complement the domain-specific treatments of SSL, which recently started appearing in application monographs [1,2]. In my opinion, this book will be an authoritative reference in the field for several years to come, at least until the field matures and a comprehensive monograph takes its place. Until then, the book is bound to have a profound effect on the directions and focus of research in semisupervised learning and machine learning in general. Online Computing Reviews Service