Chapter 4. Building Good Training Sets – Data Preprocessing

The quality of the data and the amount of useful information that it contains are key factors that determine how well a machine learning algorithm can learn. Therefore, it is absolutely critical that we make sure to examine and preprocess a dataset before we feed it to a learning algorithm. In this chapter, we will discuss the essential data preprocessing techniques that will help us to build good machine learning models.

The topics that we will cover in this chapter are as follows:

• Removing and imputing missing values from the dataset

• Getting categorical data into shape for machine learning algorithms

• Selecting relevant features for the model construction

Dealing with missing data

It is not uncommon in real-world applications that our samples are missing one or more values for various reasons. There could have been an error in the data collection process, certain measurements are not applicable, particular fields could have been simply left blank in a survey, for example. We typically see missing values as the blank spaces in our data table or as placeholder strings such as NaN (Not A Number).

Unfortunately, most computational tools are unable to handle such missing values or would produce unpredictable results if we simply ignored them. Therefore, it is crucial that we take care of those missing values before we proceed with further analyses. But before we discuss several techniques for dealing with missing values, let's create a simple example data frame from a CSV (comma-separated values) file to get a better grasp of the problem:

>>> import pandas as pd
>>> from io import StringIO
>>> csv_data = '''A,B,C,D
....

Handling categorical data

So far, we have only been working with numerical values. However, it is not uncommon that real-world datasets contain one or more categorical feature columns. When we are talking about categorical data, we have to further distinguish between nominal and ordinal features. Ordinal features can be understood as categorical values that can be sorted or ordered. For example, T-shirt size would be an ordinal feature, because we can define an order XL > L > M. In contrast, nominal features don't imply any order and, to continue with the previous example, we could think of T-shirt color as a nominal feature since it typically doesn't make sense to say that, for example, red is larger than blue.

Before we explore different techniques to handle such categorical data, let's create a new data frame to illustrate the problem:

>>> import pandas as pd
>>> df = pd.DataFrame([
...            ['green', 'M', 10.1, 'class1'], 
...            ['red', 'L', 13.5...

Partitioning a dataset in training and test sets

We briefly introduced the concept of partitioning a dataset into separate datasets for training and testing in Chapter 1, Giving Computers the Ability to Learn from Data, and Chapter 3, A Tour of Machine Learning Classifiers Using Scikit-learn. Remember that the test set can be understood as the ultimate test of our model before we let it loose on the real world. In this section, we will prepare a new dataset, the Wine dataset. After we have preprocessed the dataset, we will explore different techniques for feature selection to reduce the dimensionality of a dataset.

The Wine dataset is another open-source dataset that is available from the UCI machine learning repository (https://archive.ics.uci.edu/ml/datasets/Wine); it consists of 178 wine samples with 13 features describing their different chemical properties.

Using the pandas library, we will directly read in the open source Wine dataset from the UCI machine learning repository:

>>...

Bringing features onto the same scale

Feature scaling is a crucial step in our preprocessing pipeline that can easily be forgotten. Decision trees and random forests are one of the very few machine learning algorithms where we don't need to worry about feature scaling. However, the majority of machine learning and optimization algorithms behave much better if features are on the same scale, as we saw in Chapter 2, Training Machine Learning Algorithms for Classification, when we implemented the gradient descent optimization algorithm.

The importance of feature scaling can be illustrated by a simple example. Let's assume that we have two features where one feature is measured on a scale from 1 to 10 and the second feature is measured on a scale from 1 to 100,000. When we think of the squared error function in Adaline in Chapter 2, Training Machine Learning Algorithms for Classification, it is intuitive to say that the algorithm will mostly be busy optimizing the weights according to the larger...

Selecting meaningful features

If we notice that a model performs much better on a training dataset than on the test dataset, this observation is a strong indicator for overfitting. Overfitting means that model fits the parameters too closely to the particular observations in the training dataset but does not generalize well to real data—we say that the model has a high variance. A reason for overfitting is that our model is too complex for the given training data and common solutions to reduce the generalization error are listed as follows:

• Collect more training data

• Introduce a penalty for complexity via regularization

• Choose a simpler model with fewer parameters

• Reduce the dimensionality of the data

Collecting more training data is often not applicable. In the next chapter, we will learn about a useful technique to check whether more training data is helpful at all. In the following sections and subsections, we will look at common ways to reduce overfitting by regularization and dimensionality...

Assessing feature importance with random forests

In the previous sections, you learned how to use L1 regularization to zero out irrelevant features via logistic regression and use the SBS algorithm for feature selection. Another useful approach to select relevant features from a dataset is to use a random forest, an ensemble technique that we introduced in Chapter 3, A Tour of Machine Learning Classifiers Using Scikit-learn. Using a random forest, we can measure feature importance as the averaged impurity decrease computed from all decision trees in the forest without making any assumptions whether our data is linearly separable or not. Conveniently, the random forest implementation in scikit-learn already collects feature importances for us so that we can access them via the feature_importances_ attribute after fitting a RandomForestClassifier. By executing the following code, we will now train a forest of 10,000 trees on the Wine dataset and rank the 13 features by their respective importance...

Summary

We started this chapter by looking at useful techniques to make sure that we handle missing data correctly. Before we feed data to a machine learning algorithm, we also have to make sure that we encode categorical variables correctly, and we have seen how we can map ordinal and nominal features values to integer representations.

Moreover, we briefly discussed L1 regularization, which can help us to avoid overfitting by reducing the complexity of a model. As an alternative approach for removing irrelevant features, we used a sequential feature selection algorithm to select meaningful features from a dataset.

In the next chapter, you will learn about yet another useful approach to dimensionality reduction: feature extraction. It allows us to compress features onto a lower dimensional subspace rather than removing features entirely as in feature selection.