Test-Driven Machine Learning Control your machine learning algorithms using test-driven development to achieve quantifiable milestones Justin Bozonier BIRMINGHAM - MUMBAI Test-Driven Machine Learning Copyright © 2015 Packt Publishing All rights reserved. No part of this book may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, without the prior written permission of the publisher, except in the case of brief quotations embedded in critical articles or reviews. Every effort has been made in the preparation of this book to ensure the accuracy of the information presented. However, the information contained in this book is sold without warranty, either express or implied. Neither the author, nor Packt Publishing, and its dealers and distributors will be held liable for any damages caused or alleged to be caused directly or indirectly by this book. Packt Publishing has endeavored to provide trademark information about all of the companies and products mentioned in this book by the appropriate use of capitals. However, Packt Publishing cannot guarantee the accuracy of this information. First published: November 2015 Production reference: 1231115 Published by Packt Publishing Ltd. Livery Place 35 Livery Street Birmingham B3 2PB, UK. ISBN 978-1-78439-908-5 www.packtpub.com Credits Author Project Coordinator Justin Bozonier Paushali Desai Reviewers Proofreader Lars Marius Garshol Safis Editing Alexey Grigorev Indexer Tejal Daruwale Soni Commissioning Editor Dipika Gaonkar Graphics Jason Monteiro Acquisition Editors Divya Poojari Llewellyn Rozario Production Coordinator Melwyn Dsa Content Development Editor Nikhil Potdukhe Cover Work Melwyn Dsa Technical Editors Rupali R. Shrawane Copy Editor Yesha Gangani About the Author Justin Bozonier is a data scientist living in Chicago. He is currently a Senior Data Scientist at GrubHub. He has led the development of their custom analytics platform and also led the development of their first real time split test analysis platform which utilized Bayesian Statistics. In addition he has developed machine learning models for data mining as well as for prototyping product enhancements. Justin's software development expertise has earned him acknowledgements in the books Parallel Programming with Microsoft® .NET as well as Flow-Based Programming, Second Edition. He has also taught a workshop at PyData titled Simplified Statistics through Simulation. His previous work experience includes being an Actuarial Systems Developer at Milliman, Inc., contracting as a Software Development Engineer II at Microsoft, and working as a Sr. Data Analyst and Lead Developer at Cheezburger Network amongst other experience. Savannah Bozonier—the best partner I've ever had in life. Time and again she has made room in her life so I can push myself to do things that take an immense amount of time. Things like writing this book. My friends and colleagues for their support and help which culminated in this book: Tom Hayden, Drew Fustin, and Andrew Slotnick. My mentors across the years—Chad Boyer, Kelly Leahy, Robert Ream, James Thigpen, and Loren Bast. My parents—I don't know what it's like to be told I can't do something. My life reflects that in every way. About the Reviewers Lars Marius Garshol has worked as a consultant, product developer, and open source developer for two decades. He added Unicode support to the Opera web browser, edited a number of ISO standards, and developed the query language tolog. Later, he worked as an enterprise architect and an R&D developer. He is the developer of Duke, an open source tool for identifying near-duplicate database records. He wrote Definitive XML Application Development, published in 2002. Currently he is a software engineer at Schibsted Products & Technology in Oslo, Norway. He's working on a book on Norwegian farmhouse ale. Alexey Grigorev is an experienced software developer and data scientist with five years of professional experience. In his day-to-day job, he actively uses R and Python for data cleaning, data analysis, and modeling. He believes that testing is not only an integral part of software development, but it is also very useful for building machine learning models. www.PacktPub.com Support files, eBooks, discount offers, and more For support files and downloads related to your book, please visit www.PacktPub.com. Did you know that Packt offers eBook versions of every book published, with PDF and ePub files available? You can upgrade to the eBook version at www.PacktPub. com and as a print book customer, you are entitled to a discount on the eBook copy. Get in touch with us at [email protected] for more details. At www.PacktPub.com, you can also read a collection of free technical articles, sign up for a range of free newsletters and receive exclusive discounts and offers on Packt books and eBooks. TM https://www2.packtpub.com/books/subscription/packtlib Do you need instant solutions to your IT questions? PacktLib is Packt's online digital book library. Here, you can search, access, and read Packt's entire library of books. Why subscribe? • Fully searchable across every book published by Packt • Copy and paste, print, and bookmark content • On demand and accessible via a web browser Free access for Packt account holders If you have an account with Packt at www.PacktPub.com, you can use this to access PacktLib today and view 9 entirely free books. Simply use your login credentials for immediate access. Table of Contents Preface v Chapter 1: Introducing Test-Driven Machine Learning 1 Test-driven development 2 The TDD cycle 3 Red 4 Green 4 Refactor 4 Behavior-driven development 5 Our first test 6 The anatomy of a test 8 Given 9 When 9 Then 9 TDD applied to machine learning 10 Dealing with randomness 11 Different approaches to validating the improved models 16 Classification overview 16 Regression 16 Clustering 17 Quantifying the classification models 17 Summary 19 Chapter 2: Perceptively Testing a Perceptron 21 Getting started 21 Summary 33 [ i ] Table of Contents Chapter 3: Exploring the Unknown with Multi-armed Bandits 35 Understanding a bandit 35 Testing with simulation 36 Starting from scratch 37 Simulating real world situations 40 A randomized probability matching algorithm 43 A bootstrapping bandit 45 The problem with straight bootstrapping 48 Multi-armed armed bandit throw down 50 Summary 52 Chapter 4: Predicting Values with Regression 53 Refresher on advanced regression 53 Regression assumptions 54 Quantifying model quality 54 Generating our own data 55 Building the foundations of our model 56 Cross-validating our model 64 Generating data 65 Summary 66 Chapter 5: Making Decisions Black and White with Logistic Regression 67 Generating logistic data 68 Measuring model accuracy 70 Generating a more complex example 71 Test driving our model 72 Summary 77 Chapter 6: You're So Naïve, Bayes 79 Gaussian classification by hand 79 Beginning the development 81 Summary 99 Chapter 7: Optimizing by Choosing a New Algorithm 101 Upgrading the classifier 101 Applying our classifier 106 Upgrading to Random Forest 111 Summary 115 [ ii ] Table of Contents Chapter 8: Exploring scikit-learn Test First 117 Test-driven design 118 Planning our journey 118 Creating a classifier chooser (it needs to run tests to evaluate classifier performance) 119 Getting choosey 131 Developing testable documentation 134 Decision trees 134 Summary 137 Chapter 9: Bringing It All Together 139 Starting at the highest level 141 The real world 158 What we've accomplished 165 Summary 165 Index 167 [ iii ]