Browse Books

Reviewer: Jose Hernandez-Orallo

Two of the ever-burning issues in machine learning are the ability to learn from complex data and the ability to adequately handle background knowledge. Both issues are closely related since most regularities in complex data can only be disentangled in a reasonable time by using early features and concepts that can pave the way for future discoveries. Examples of this can be found every day in human learning and scientific discovery. From time to time, a group of researchers in machine learning and other areas come together to give a new boost to existing issues. While in the late 1980s and 1990s it was the field of inductive logic programming that captained this research, in recent years researchers from new and old areas are converging toward new names and views. Statistical relational learning is one of these new terms, along with other related (or almost synonymous) terms such as multirelational data mining, probabilistic logic learning, and learning from structured data. While the word "relational" represents the "target," namely, relational or complex data, the word "statistical" does not refer to "statistics" in a strict sense, but to techniques that place probability and related concepts at the first plane (stochastic processes, Bayesian methods, and Markov models). This book presents an up-to-date and comprehensive selection of topics in statistical relational learning. The editors are two of the researchers who have most actively participated in the move toward this new confluence, playing an important role in meetings and activities, with the goal of shaping a new field. As a result, they have been able to gather an excellent collection of chapters from top-level authors in the field of machine learning. After the introduction, the following 19 chapters start with fundamental or instrumental topics such as graphical models, inductive logic programming, and conditional random fields. Next, the bulk of the book (chapters 5 to 12) is properly devoted to approaches that combine probabilistic languages or inference techniques on relational data, some of them "frame-based" and some of them logic based. From chapter 13 to the end of the book, the topics are more specific, covering particular systems, languages, tasks, application areas, and machine-learning paradigms. As usually happens with collections, some chapters are clearly more introductory than others, but there are many examples in most chapters. While the content is, in general, of excellent quality, the book's organization might have been improved. Although the order for chapters is sensible, chapters frequently overlap, and the authors do not always mention this fact. For instance, chapter 4 devotes several pages to graphical models, while chapter 2 is completely devoted to graphical models. Of course, it is a good thing that each chapter can be read independently from the rest, but only in a very few chapters can one find short references to or comments about other chapters, even though the possible and useful connections between chapters are numerous. The introduction (chapter 1) does not fill this gap, and gives no hints on whether the book's chapters should be read sequentially or independently. Additionally, the introduction of the term "statistical relational learning" and its history seems to be defined for researchers who are already in the area, not for newcomers. In fact, a short section such as "The Need for a Unifying Framework," given in chapter 12, would have provided a more direct and clear invitation to the field. A more cohesive structure would have given the book more points to be appropriate not only for "researchers," but also for "graduate students," as the editors claim. Graduate students will clearly need assistance in choosing the right chapters, and in ascertaining which previous knowledge on statistics, machine learning, programming languages, databases, and computer science in general is required for each chapter. On the other hand, if this book is intended for researchers in machine learning and data mining who would like to enter this new field, to acquire a good state-of-the-art knowledge, or to use it as a reference, then the organization and presentation is sober but sufficient. In any case, the book clearly fills the need for an integrated and up-to-date introduction to a very important area (either called "statistical relational learning" or referred to in other terms); it is an area that will receive more and more attention in the future. The presence of some extremely excellent contributors is a credit to the book. Additionally, in many cases, these authors have presented their more mature ideas and works, which will make this book highly referenced, and not as prone to become obsolete in the near future. Online Computing Reviews Service