Feature Engineering and Selection (Chapman & Hall/CRC Data Science Series) 1st Edition

Interpretable IA is a must. Good feature engineering allows interpretable models to achieve black box models' performance. This book contains a deep research on those techniques; the book also addresses interesting turn arounds to prevent overfitting. A must read for any data scientist.

I read Kuhn and Johnson's Applied Predictive Modeling text when it first came out and have the upmost respect for their ability to explain difficult topics with an applied slant. Extremely good book. This text follows very much in the same vein, and I do recommend it, especially for data science practitioners with less than 5 years experience. As someone who has been working as a data scientist for nearly a decade, I was even able to glean some feature engineering nuggets that I had not encountered before (and had ashamedly forgotten).Now ... here is my critique:First, Kuhn and Johnson should be absolutely infuriated with whoever at CRC did the layout for this book. The plot/graphic decision placement in the hardcopy version of the text is just bad ... and there are some GREAT plots in this book. It feels like only about 25% of the time a plot/graphic actually resides on the same page(s) as which it is referenced. This requires the reader to constantly be flipping back and forth between pages disrupting the flow of what Kuhn and Johnson are attempting to present. In other words, the ergonomics of this book don't work, and by halfway through my read, I pretty much wanted to stop reading. Maybe the digital version of the book is better, but the hardcopy just had this issue that was very distracting for me. I never thought I'd write a review that included graphic placement as a topic, but this text warrants it.Second, the book lacks a comprehensive treatment of feature selection with time series data. While temporal data is addressed in the first half of the book. it is not addressed in the last three chapters regarding feature selection (which may be among the most valuable and content-rich in the entire book). Their suggestion of external and internal resampling with cross-fold validation is extremely valuable when using IID data. However, they make no recommendation or even mention about how to utilize this resampling ideology with temporal data.Don't get me wrong ... I really, really like this book and I do think there is incredibly valuable content that nearly every data scientist (experienced or new) would be able to take away from it. However, I just can't give it a glowing 5-star review when it feels like there is a bit of a void when it comes to addressing feature selection involving time series data.

Although this book provides a holistic view of practical methods to handle features, it really needs to expand the explanation on each method, such as thorough examples, code snippets instead of an index of book recommendations if you want to know each method furthermore.

The content is superb, and overall, no regrets on the purchase.I wish it had full Kindle support, though, so it would be more comfortable to read with custom font sizes, instead of zooming and panning.

Given that I think that Applied Predictive Modeling (APM) is one of the best applied statistics books ever written, I had high hopes for this and I am not disappointed. The beginning of this book recaps many of the key ideas from APM and then takes you through the process of choosing and tweaking the features you will want to use for modeling. This book shares the same easy to read style as APM and it has excellent well motivated advice. I have been doing statistics and statistical programming for decades and I still found many novel data manipulation tricks, graphics and methods.The book itself does not directly integrate and discuss code (instead there is pseudocode to explain ideas/algorithms). Happily the github repository has the materials to reproduce the work. While I would have preferred the authors directly cover coding or provided a full Bookdown version of the manuscript, an intermediate R programmer who is comfortable with tidyverse will be able to untangle most everything from the repo. Grab a copy and be impressed.