Conformal prediction is based on a family of measures that are based on a combination of algorithmic information theory and machine learning techniques, providing a means for reliable estimation with consideration to confidence values and with applications to classification, regression, and clustering. The framework can be adapted to traditional machine learning algorithms such as support vector machines, k-nearest neighbors, and neural networks, to mention but a few. This book presents the theory, the extensions of the methods when translated to classical machine learning frameworks, and finally applications to network traffic, computational medicine, and more.

I found it useful to actually start at the end of the book, from the application to the theory, given that the theory is non-conventional and far from mainstream, even to me as a researcher in algorithmic complexity. Indeed, “Applications,” Part 3, has self-contained chapters that explain the basic framework in a more accessible way before going to Part 1, “Theory,” and Part 2, “Adaptations.”

The only issue I have with this book is that it does not walk the reader through the steps necessary to build an intuition of the measure of conformal prediction introduced. For example, it lacks examples in the theory part; even though one may argue that the examples are in the applications part of the book, I think the authors could have been more sympathetic to the reader and helped him build a better and quicker intuition of how the measure is really applied, how it works, and why it works (beyond the very specific and isolated example of the spam filter in chapter 2). That is why I suggest reading this book from back to front; however, even the applications part lacks details on the implementation and numerical calculation of the applied measures.

In summary, this is the only general textbook I’ve seen on this promising topic. Though more could have been done to make the topic more accessible from the point of view of providing more context--a survey of the field in which it is inserted, and being more sympathetic toward building intuition and providing step-by-step calculations to make it less cryptic and help the adoption of its proposed techniques--the book is highly recommended for people looking for formal machine learning techniques that can guarantee theoretical soundness and reliability.