Imagine an imbalanced dataset of cancer patients with highly skewed data having only 0.01 percent positive cancer cases. A naive or dumb machine that calls out “no cancer” to all queries would appear to be 99.99 percent accurate, and could even be misconstrued as a good prediction model over a competing machine learning algorithm. Additionally, the consequences of such a misclassification could be disastrous for patients with cancer. A comprehensive knowledge of machine learning, therefore, would be incomplete without a fair understanding of such predicaments and how to resolve them.

This book promises to engage the reader by providing a vivid picture of the problems associated with imbalanced datasets, specific aspects and approaches to solve the problems, and assessment metrics. The narrative is ordered and easy to understand.

A dozen authors contribute to the book’s eight chapters: “Introduction,” “Foundations of Imbalanced Learning,” “Imbalanced Datasets: From Sampling to Classifiers,” “Ensemble Methods for Class Imbalance Learning,” “Class Imbalance Learning Methods for Support Vector Machines,” “Class Imbalance and Active Learning,” “Nonstationary Stream Data Learning with Imbalanced Class Distribution,” and “Assessment Metrics for Imbalanced Learning.” There aren’t any competing books on imbalanced learning.

Leaving aside the usual issues associated with multiple contributors--for example, the high chance for repetition or the difficulty of maintaining uniformity in presentation style--the book does justice to machine learning by bringing out issues related to imbalanced datasets. The significance of precision and recall is introduced or explained in multiple chapters, but presented as it is from varying perspectives, it doesn’t affect the interest of the reader.

The editors have succeeded in maintaining coherency and consistency while presenting content. For instance, while the terms F-score or F1 score could also have been used, the consistent use of F-measure throughout the book is noteworthy. Consistency is also visible in the illustrations involving precision and recall. With different authors assigned to different chapters, it is extremely difficult to trace out errors. Only one instance was detected: welding flaw was introduced as an example for imbalanced datasets in chapter 2. The case associated with welding flaws is a more apt example for discussions on anomaly detection and outliers. Anomalies, by definition, do not constitute a class or cluster by themselves, even if skewness is present as an attribute on the data.

This book certainly qualifies as a reference for graduate studies in machine learning. Research students are sure to find it highly valuable and a prized possession, especially taking into account the wealth of supporting literature that the authors have brought to the fore.