Book Image

Data Cleaning and Exploration with Machine Learning

By : Michael Walker
Book Image

Data Cleaning and Exploration with Machine Learning

By: Michael Walker

Overview of this book

Many individuals who know how to run machine learning algorithms do not have a good sense of the statistical assumptions they make and how to match the properties of the data to the algorithm for the best results. As you start with this book, models are carefully chosen to help you grasp the underlying data, including in-feature importance and correlation, and the distribution of features and targets. The first two parts of the book introduce you to techniques for preparing data for ML algorithms, without being bashful about using some ML techniques for data cleaning, including anomaly detection and feature selection. The book then helps you apply that knowledge to a wide variety of ML tasks. You’ll gain an understanding of popular supervised and unsupervised algorithms, how to prepare data for them, and how to evaluate them. Next, you’ll build models and understand the relationships in your data, as well as perform cleaning and exploration tasks with that data. You’ll make quick progress in studying the distribution of variables, identifying anomalies, and examining bivariate relationships, as you focus more on the accuracy of predictions in this book. By the end of this book, you’ll be able to deal with complex data problems using unsupervised ML algorithms like principal component analysis and k-means clustering.
Table of Contents (23 chapters)
1
Section 1 – Data Cleaning and Machine Learning Algorithms
5
Section 2 – Preprocessing, Feature Selection, and Sampling
9
Section 3 – Modeling Continuous Targets with Supervised Learning
13
Section 4 – Modeling Dichotomous and Multiclass Targets with Supervised Learning
19
Section 5 – Clustering and Dimensionality Reduction with Unsupervised Learning

Summary

This chapter explored how to use decision trees for classification problems. Although the examples in this chapter all involved a binary target, the algorithms we worked with can also handle multiclass problems. Unlike the switch from logistic to multinomial logistic regression, few changes need to be made to use the algorithms well when our target has more than two values.

We looked at two approaches to dealing with the high variance of decision trees. One approach is to use a random forest, which is a form of bagging. This will reduce the variance in our predictions. Another approach is to use gradient-boosted decision trees. Boosting can help us capture very complicated relationships in the data, but there is a non-trivial risk of overfitting. It is particularly important to tune our hyperparameters with that in mind.

In the next chapter, we explore another well-known algorithm for classification: K-nearest neighbors.