Hands-On Ensemble Learning with R

Hands-On Ensemble Learning with R

By : Prabhanjan Narayanachar Tattar

Buy this Book

Hands-On Ensemble Learning with R

By: Prabhanjan Narayanachar Tattar

Buy this Book

Overview of this book

Ensemble techniques are used for combining two or more similar or dissimilar machine learning algorithms to create a stronger model. Such a model delivers superior prediction power and can give your datasets a boost in accuracy. Hands-On Ensemble Learning with R begins with the important statistical resampling methods. You will then walk through the central trilogy of ensemble techniques – bagging, random forest, and boosting – then you'll learn how they can be used to provide greater accuracy on large datasets using popular R packages. You will learn how to combine model predictions using different machine learning algorithms to build ensemble models. In addition to this, you will explore how to improve the performance of your ensemble models. By the end of this book, you will have learned how machine learning algorithms can be combined to reduce common problems and build simple efficient ensemble models with the help of real-world examples.

Hands-On Ensemble Learning with R

Contributors

Preface

Free Chapter

Introduction to Ensemble Techniques

Datasets

Statistical/machine learning models

The right model dilemma!

An ensemble purview

Complementary statistical tests

Summary

Bootstrapping

Technical requirements

The jackknife technique

Bootstrap – a statistical method

The boot package

Bootstrap and testing hypotheses

Bootstrapping regression models

Bootstrapping survival models*

Bootstrapping time series models*

Summary

Bagging

Technical requirements

Classification trees and pruning

Bagging

k-NN classifier

k-NN bagging

Summary

Random Forests

Technical requirements

Random Forests

Variable importance

Proximity plots

Random Forest nuances

Comparisons with bagging

Missing data imputation

Clustering with Random Forest

Summary

The Bare Bones Boosting Algorithms

Technical requirements

The general boosting algorithm

Adaptive boosting

Gradient boosting

Using the adabag and gbm packages

Variable importance

Comparing bagging, random forests, and boosting

Summary

Boosting Refinements

Technical requirements

Why does boosting work?

The gbm package

The xgboost package

The h2o package

Summary

The General Ensemble Technique

Technical requirements

Why does ensembling work?

Ensembling by voting

Ensembling by averaging

Stack ensembling

Summary

Ensemble Diagnostics

Technical requirements

What is ensemble diagnostics?

Ensemble diversity

Pairwise measure

Interrating agreement

Summary

Ensembling Regression Models

Technical requirements

Pre-processing the housing data

Visualization and variable reduction

Regression models

Bagging and Random Forests

Boosting regression models

Stacking methods for regression models

Summary

Ensembling Survival Models

Core concepts of survival analysis

Nonparametric inference

Regression models – parametric and Cox proportional hazards models

Survival tree

Ensemble survival models

Summary

Ensembling Time Series Models

Technical requirements

Time series datasets

Time series visualization

Core concepts and metrics

Essential time series models

Bagging and time series

Ensemble time series models

Summary

What's Next?

Bibliography

References

R package references

Index

Customer Reviews

5 star

4 star

3 star

2 star

1 star

Nonparametric inference

Survival data is subject to censoring and we need to introduce a new quantity to capture this information. Suppose that we have a n IID random sample of lifetime random variables in , and we know that the event of interest might have occurred or that it will occur sometime in the future. The additional information is captured by the Kronecker indicator variable, :

Thus, we have n pairs of random observations in the Ts and s, . To obtain the estimates of the cumulative hazard function and the survival function, we will need an additional notation. Let denote the unique times of Ts at which the event of interest is observed. Next, we denote to represent the number of observations that are at risk just before times and the number of events that occur at that time. Using these quantities, we now propose to estimate the cumulative hazard function using the following: