Applied Machine Learning – Jason Hodson (1)

When tuning hyperparameters to balance performance and overfitting, we’ll still want to consider n_estimators, max_depth, and min_samples_split. However, now we’ll also introduce two new hyperparameters: learning_rate and subsample. The Problem of Predictive Power As mentioned earlier, more predictive power comes with more problems. In this specific dataset example, you likely wouldn’t use a GBM, as it’s too powerful a model. For the sake of the example and continuity with the decision tree and random forest sections, we’ll continue to use the same data.

You’ll see the performance impacts of the GBM later on in our use cases, so focus on learning the hyperparameters and understanding their impact as you go through this simpler example. The learning_rate hyperparameter adjusts how much each tree contributes to the final model. The higher the number, the faster the model runs because learning is sped up—but then you risk overfitting. The risk of overfitting is reduced with a lower value. The lowest value is technically 0, and there is no upper limit on what you can input.

Think about raising or lowering the learning rate as a trade-off in how quickly the model makes an assumption based on the data it sees. A high learning rate means the model is more willing to trust the data it sees right away. A lower learning rate makes the model more skeptical of the data you give it. This runs the risk that your model never learns new information that may be valuable. The subsample hyperparameter introduces randomness into each iteration of your model, which is another way to reduce overfitting.

This number represents a percentage of the rows that each iteration of the model can use, so it’s less abstract than the learning_rate hyperparameter. Let’s start with some simple adjustments to these new hyperparameters to see how the results are impacted, as shown in Listing 5.86.

Production E-Book Kelly O’Callaghan Typesetting E-Book III-satz, Germany We hope that you liked this e-book. Please share your feedback with us and read the Service Pages to find out how to contact us. The Library of Congress Cataloging-in-Publication Control Number for the printed edition is as follows: 2025053188 ISBN 978-1-4932-2758-7 (print) ISBN 978-1-4932-2759-4 (e-book) ISBN 978-1-4932-2760-0 (print and e-book) © 2026 by Rheinwerk Publishing Inc., Boston (MA) 1st edition 2026 1kitap1.com/en Notes on Usage This e-book is protected by copyright.

By purchasing this e-book, you have agreed to accept and adhere to the copyrights. You are entitled to use this e-book for personal purposes. You may print and copy it, too, but also only for personal use. Sharing an electronic or printed copy with others, however, is not permitted, neither as a whole nor in parts. Of course, making them available on the internet or in a company network is illegal as well.

For detailed and legally binding usage conditions, please refer to the section Legal Notes. Notes on the Screen Presentation You are reading this e-book in a file format (EPUB or Mobi) that makes the book content adaptable to the display options of your reading device and to your personal needs.

That’s a great thing; but unfortunately not every device displays the content in the same way and the rendering of features such as pictures and tables or hyphenation can lead to difficulties. This e-book was optimized for the presentation on as many common reading devices as possible. If you want to zoom in on a figure (especially in iBooks on the iPad), tap the respective figure once. By tapping once again, you return to the previous screen.

You can find more recommendations on the customization of the screen layout on the Service Pages. 1kitap1.com/en Table of Contents Notes on Usage Table of Contents Preface 1 Introduction 1.1 Aligning on Nomenclature 1.2 Learning to Google (or Prompt) 1.2.1 What Can You Find with Google?

This is a short excerpt from the opening of “” by Unknown, quoted for review and introduction purposes. All rights belong to the copyright holders.

Book Information

• Unique ID: f2fa38988b468d14
• File Extension: .pdf
• File Size: 21,252,411 bytes (20.268 MB)
• Title: –
• Author: Unknown
• ISBN: 9781493227587, 9781493227594, 9781493227600
• Pages: 829
• Language: English (en)

Reading & Word Statistics

• Estimated Reading Time: 593.41 minutes
• Total Words: 118,683
• Total Characters: 758,807
• Average Words per Page: 143.16
• Average Characters per Page: 915.33

Most Frequent Words

data (1648), model (1559), section (844), test (775), use (683), listing (628), figure (555), train (507), columns (442), column (433), create (421), code (406), date (404), values (400), print (399), we’ll (395), case (364), accuracy (358), it’s (348), see (340), learning (332), using (293), split (289), results (287), time (281), you’re (280), number (268), feature (267), function (267), first (248), one (245), customer (242), com (241), import (236), random (234), dataset (230), need (229), predict (228), target (227), kitap (226), models (226), predictions (224), order (224), process (222), new (221), mean (218), machine (215), score (210), pred (210), grid (207), we’re (203), day (200), example (199), shown (197), search (194), crime (192), like (192), value (191), decision (190), regression (187), you’ll (184), now (183), get (178), best (175), let’s (174), mae (166), fit (165), dummy (163), important (161), error (160), sklearn (157), also (154), tree (154), approach (153), auc (153), training (152), however (151), count (151), make (145), chapter (145), plot (145), max (144), next (141), dataframe (141), name (139), week (139), company (138), used (138), variable (136), many (135), business (134), file (134), last (134), treatment (129), state (128), don’t (127), control (125), book (125), area (125), description (125).

PDF Download