Applied Predictive Modeling

tl;dr: A brilliant book covering Predictive modelling in R. With a strong practical bent it walks the reader through the application of modern classification and regression techniques to a broad number of varied and interesting data sets. It uses existing packages where possible so you can jump straight in (great for Kagglers) but there is a lot here to master. It is especially strong on preprocessing (both unsupervised and supervised), model tuning and model assessment. Should not be your first book on R or data analytics but the best balance of Practical application without foregoing theory that I have seen. It is wonderful to see how professional data analysts approach predictive modelling tasks. The data sets are not toy models to highlight approaches but interesting and complex problems from a wide variety of disciplines.(Note that this book does not cover Time Series, Generalised Additive Models and Ensemble's of different models).Review:Data science has become very popular due to the increase in computing power (including things like AWS), the amount of data that is accessible on the internet and a number of open-source tools (R and Python for example) that allow even relative beginners to complete quite sophisticated models. Coursera allows for one to complete courses on Machine Learning for free and sites like Kaggle have even turned it into something of a sport where people compete to create predictive models for money or even job interviews. Part of the excitement is that Predictive models can be applied to almost any field you can think of.Given the easy access to predict things using sophisticated techniques, the number of books on machine learning, data mining and predictive data analytics has grown to fill the demand of people looking to learn about the field. As data science is itself a combination of many different disciplines (statistics, computer science, artificial intelligence etc) there are many different points of entry. For this reason books can often be placed on a spectrum from straightforward examples of already constructed programs to theoretical textbooks with lots of mathematical background and constructing approaches from scratch. "Applied Predictive Modeling" tries to find a middle ground between these two approaches though it unashamedly sides with the practical. In contrast to many other works though, it utilises existing packages (notably caret) rather than having the reader construct the approaches themselves in code.Applied Predictive Modeling contains 20 Chapters set out to be quasi-independent whilst still being a coherent book. An abstract opens each chapter followed by sections discussing the approaches used. The writing is excellent, very easy to follow and wonderfully informative with an excellent choice of example data sets. The discussions are not afraid to highlight the problems of different approaches - in one of the latter chapters noise is deliberately added to a data set so the differing impact can be seen on a range of models. Theory is discussed insofar as it is useful for understanding the use of certain approaches and references to further reading are clearly given. The chapters conclude with a summary before containing a computation section which contains all of the R code used for the chapters with some discussion where important. Finally most chapters have a section containing exercises. Usefully these exercises use different data sets so are not merely regurgitation of what one has just read. The chapters also have independent Bibliographies which is a little annoying when reading the book cover to cover, but makes it excellent as a reference book.After a few chapters of overview the chapters largely work through the components in the process of Data Analytics; data-splitting, pre-processing chapters cover transforming, centering, dealing with missing values and setting up the data for the application of models. The next section of the book covers Regression models. It utilises a Pharmaceutical dataset and works through the creation of models of increasing complexity. A chapter then works through an example of concrete strength prediction based on ingredients to show clearly how regression applications work end to end. A number of chapters then look at classification algorithms using the construction of a data model from a kaggle competition from late 2010 on University Grants. This highlights what this book offers that I have not seen in other comparable books - real life examples on the steps a professional analyst takes in the construction of a model. The reader is almost always watching the construction of a real model throughout the discussion of the differing approaches. The book does discuss theory where it is useful. But rather then going into the miniature of constructing things directly in code to highlight the underlying structure, existing packages are used where possible. This lowers the barrier to getting started on using the techniques. Finally the book is rounded out with chapters on model tuning, detecting variable importance, how to handle class imbalances and some broader issues in modelling all again using real data sets from different fields.The authors have created an R package for the book containing the code and data sets used and an excellent website and blog. The book ranges broadly across disciplines and includes separate data sets for the exercises, in all I count 21 data sets ranging from concrete strength to caravan insurance that are either covered in the book or are given as exercises in the chapters.In short I congratulate the authors on an excellent book that I look forward to working through in depth over the coming months. If you are looking to improve your predictive modelling and are short of professional standard, this is the book you are looking for. Whilst there are loads to learn and master - you can jump in and use things from the book very quickly thanks to its use of impressive packages. One area I would love to see added to future editions would be the ensembling of different models.

There are many fine math-oriented predictive modeling books, such as Hastie (The Elements of Statistical Learning: Data Mining, Inference, and Prediction, Second Edition (Springer Series in Statistics)). Kuhn et al consider them "sister texts" and begin immediately to differentiate-- their approach is hands on and practical, for the express purpose of demonstrating HOW to sort, structure and predict via Python or R, for the purpose of accuracy and understanding of the DATA and trends, NOT learning the underlying math.For a couple of pharmaceutical guys, (who BTW use R extensively, I've been an analyst in that industry), you'd think the examples would be new chemical or biological entities. Not so! The cases are fun and exciting, ranging from the nontrivial compression strength of concrete (want that bridge to hold when you cross?) to fuel economy, credit scoring, success in grant applications (boy their colleagues will love that one!), and cognitive impairment. I evaluate technology for patents at payroy dot com, and we have a log likelihood model using Bayesian and Monte Carlo that their grant section helped translate seamlessly to R! We're NOT talking pie in the sky pseudo code here, but real life, real results recipes.The authors talk about the "scholarly veil" -- meaning we general workers and researchers don't always "deserve" to see the underlying process, software and data (and, other than open source, often can't afford it). Wow, do they pop that myth! These authors are relentless in giving every detail, from design and binning to sorting and stacking to ANOVA, regressions, trees, error methods-- the whole ball of wax with live data and live R coding-- all on a shoestring budget! I guarantee you can start with basic stats and run a very well designed predictive model with the methods they detail, without having to pop for SAP/ IBM or SPSS.One caveat-- even though they don't assume advanced partial differential equations or even probability theory, the R code and methods are at a fast clip. I'd say they are assuming you either have, or will fill in, with R basics and practice or experience. This is NOT a "how to use R" manual, even though it is in a sense-- it is a "how to apply R correctly and robustly in a way that will pass a juried look at your methods and conclusions." Again, REAL WORLD. For comparison, I'd put the math at advanced undergrad and the R at grad level/ professional practice levels. This will make the title excellent both for learning and professional reference. At this writing, the book is hard to find, and being marked up by resellers-- a tribute to its value and demand right out of the gate.Springer is never cheap, but also never shabby-- the book is typically gorgeous, well edited, combed for errors (the code ran fine on my antique R download-- even though it's free, I'm hesitant to have to learn a new version!), and pedagogically awesome if you're considering this for a class. We recommend books for our library purchasers and of the 25 actively screened in this category (including a focus on prediction, not just data mining), this is in the top three with Hastie above! Highly recommended for research, augmentation, reference, as well as deep study. Lots of insights, too, about where big data, ML, mining and prediction are now and where they are going-- predicting prediction's future.Library Picks reviews only for the benefit of Amazon shoppers and has nothing to do with Amazon, the authors, manufacturers or publishers of the items we review. We always buy the items we review for the sake of objectivity, and although we search for gems, are not shy about trashing an item if it's a waste of time or money for Amazon shoppers. If the reviewer identifies herself, her job or her field, it is only as a point of reference to help you gauge the background and any biases.