Book Image

Machine Learning at Scale with H2O

By : Gregory Keys, David Whiting
Book Image

Machine Learning at Scale with H2O

By: Gregory Keys, David Whiting

Overview of this book

H2O is an open source, fast, and scalable machine learning framework that allows you to build models using big data and then easily productionalize them in diverse enterprise environments. Machine Learning at Scale with H2O begins with an overview of the challenges faced in building machine learning models on large enterprise systems, and then addresses how H2O helps you to overcome them. You’ll start by exploring H2O’s in-memory distributed architecture and find out how it enables you to build highly accurate and explainable models on massive datasets using your favorite ML algorithms, language, and IDE. You’ll also get to grips with the seamless integration of H2O model building and deployment with Spark using H2O Sparkling Water. You’ll then learn how to easily deploy models with H2O MOJO. Next, the book shows you how H2O Enterprise Steam handles admin configurations and user management, and then helps you to identify different stakeholder perspectives that a data scientist must understand in order to succeed in an enterprise setting. Finally, you’ll be introduced to the H2O AI Cloud platform and explore the entire machine learning life cycle using multiple advanced AI capabilities. By the end of this book, you’ll be able to build and deploy advanced, state-of-the-art machine learning models for your business needs.
Table of Contents (22 chapters)
1
Section 1 – Introduction to the H2O Machine Learning Platform for Data at Scale
5
Section 2 – Building State-of-the-Art Models on Large Data Volumes Using H2O
11
Section 3 – Deploying Your Models to Production Environments
14
Section 4 – Enterprise Stakeholder Perspectives
17
Section 5 – Broadening the View – Data to AI Applications with the H2O AI Cloud Platform

Model optimization with grid search

Choosing an algorithm for building a predictive model is not enough. Many algorithms have hyperparameters whose values have a direct impact on the predictive power of the model. So, how do you choose values for your hyperparameters?

A brute-force method would create a grid of all possible values and search over them. This approach is computationally expensive, takes an inordinate amount of time, and ultimately, yields results that are not much better than what we could achieve by other means. We have outlined a strategy for grid search that has proved effective in building optimized models while running in a reasonable amount of time.

The general strategy entails, first, tuning a few key parameters using a Cartesian grid search. These key parameters are those we expect to have the biggest impact on the results. Then, we fine-tune other parameters using a random grid search. This two-stage approach allows us to hone in on the computationally...