This is a featured recipe from the IPython Cookbook, the definitive guide to high-performance scientific computing and data science in Python.
In Statistical Data Analysis, we are interested in getting insight into data, understanding complex phenomena through partial observations, and making informed decisions in the presence of uncertainty. In Machine Learning, we are still interested in analyzing and processing data using statistical tools. However, the goal is not necessarily to understand the data, but to learn from it.
Learning from data is close to what we do, as humans. From our experience, we intuitively learn general facts and relations about the world, even if we don't fully understand its complexity. The increasing computational power of computers makes them able to learn from data, too. That's the heart of machine learning, a modern and fascinating branch of artificial intelligence, computer science, statistics, and applied mathematics.
This featured recipe is a hands-on introduction to the most fundamental concepts in machine learning. These concepts are routinely used by data scientists. We will illustrate them with scikit-learn, a popular and user-friendly Python package for machine learning.
Let's describe the fundamental definitions and concepts of machine learning.
In machine learning, most datasets can be represented as tables containing numerical values. Every row is called an observation, a sample, or a data point. Every column is called a feature or a variable.
Let's call $N$ the number of rows (or number of points), and $D$ the number of columns (or number of features). The number $D$ is also called the dimensionality of the data. The reason is that we can view this table as a set $E$ of vectors in a space with $D$ dimensions (or vector space). Here, a vector $x$ contains $D$ numbers $(x_1, ..., x_D)$, also called components. This mathematical point of view is very useful and we will use it throughout this recipe.
One generally makes the distinction between supervised learning and unsupervised learning.
Supervised learning is when we have a label $y$ associated to every data point $x$. The goal is to learn the mapping from $x$ to $y$ from our data. The data gives us this mapping for a finite set of points, but what we want is to generalize this mapping. In other words, we want to find the label of any point $x$ that does not belong to our data.
Unsupervised learning is when we don't have any labels. What we want to do is discover some hidden structure in the data.
Mathematically, supervised learning consists of finding a function $f$ that maps a set of points $E$ to a set of labels $F$, knowing a finite set of associations $(x, y)$ which is given by our data. This is what generalization is about: after observing the pairs $(x_i, y_i)$, given a new $x$, we are able to find the corresponding $y$ by applying the function $f$ to $x$.
It is a common practice to split the set of data points into two subsets: the training set and the test set. We learn the function $f$ on the training set, and test it on the test set. This is essential when assessing the predictive power of a model. By training and testing a model on the same set, our model may not be able to generalize well. This is the fundamental concept of overfitting, which we will detail later.
One generally makes the distinction between classification and regression, two particular instances of supervised learning.
The figure below illustrates the difference between classification and regression.
Broadly speaking, unsupervised learning helps us discover systemic structures in our data. It is harder to grasp than supervised learning, in that there is no precise question and answer in general.
Here are a few important terms related to unsupervised learning:
In a supervised learning context, when our data contains many features, it is sometimes necessary to choose a subset of them. The features we want to keep are those that are most relevant to our question. This is the problem of feature selection.
Additionally, we may want to extract new features by applying complex transformations on our original dataset. This is feature extraction. For example, in computer vision, training a classifier directly on the pixels is not the most efficient method in general. We may want to extract the relevant points of interest or make appropriate mathematical transformations. These steps depend on our dataset and on the questions we want to answer.
For example, it is often necessary to preprocess the data before learning models. Feature scaling (or data normalization) is a common preprocessing step where features are linearly rescaled to fit in the range $[-1,1]$ or $[0,1]$.
Feature extraction and feature selection involve a balanced combination of domain expertise, intuition, and mathematical methods. These early steps are crucial, and they may be even more important than the learning steps themselves. The reason is that the few dimensions that are relevant to our problem are generally hidden in the high dimensionality of our dataset. We need to uncover the low-dimensional structure of interest to improve the efficiency of our learning models.
Chapter 8 illustrates a few feature selection and feature extraction methods. Methods that are specific to signals, images or sounds are covered in Chapter 10 and Chapter 11. Here are a few further references:
A central notion in machine learning is the trade-off between overfitting and underfitting. A model may be able to represent our data accurately. However, if it is too accurate, it may not generalize well to unobserved data. For example, in face recognition, a too-accurate model would be unable to identity someone who styled their hair differently that day. The reason is that our model may learn irrelevant features in the training data. On the contrary, an insufficiently trained model would not generalize well either. For example, it would be unable to correctly recognize twins.
A popular solution to reduce overfitting consists of adding structure to the model, for example with regularization. This method favors simpler models during training (Occam's razor).
The bias-variance dilemma is closely related. The bias of a model quantifies how precise a model is across training sets. The variance quantifies how sensitive the model is to small changes in the training set. A robust model is not overly sensitive to small changes. The dilemma involves minimizing both bias and variance; we want a precise and robust model. Simpler models tend to be less accurate but more robust. Complex models tend to be more accurate but less robust.
The importance of this trade-off cannot be overstated. This question pervades the entire discipline of machine learning. Chapter 8 contains many examples.
As we will see in Chapter 8, there are many supervised and unsupervised algorithms. For example, well-known classifiers that we will cover include logistic regression, nearest-neighbors, Naive Bayes, support vector machines. There are many others algorithms that we couldn't cover.
No model performs uniformly better than the others. One model may perform well on one dataset and badly on another. This is the question of model selection.
We will see systematic methods for assessing the quality of a model on a particular dataset (notably cross-validation). In practice, machine learning is not an "exact science" in that it frequently involves trials and errors. We need to try different models and empirically choose the one that performs best.
That being said, understanding the details of the learning models allows us to gain intuition about which model is best adapted to our current problem.
Here are a few references on this question:
An important class of machine learning methods that we couldn't cover in Chapter 8 include neural networks and deep learning. Deep learning is the subject of very active research in machine learning. Many state-of-the-art results are currently achieved by deep learning methods.
Let's now introduce the basics of the machine learning package scikit-learn. This package is the main tool we are going to use throughout the chapter. Its clean API makes it really easy to define, train, and test models. Plus, scikit-learn is specifically designed for speed and (relatively) big data.
We will show here a very basic example of linear regression in the context of curve fitting. This toy example will allow us to illustrate key concepts such as linear models, overfitting, underfitting, regularization, and cross-validation.
You will find all instructions to install scikit-learn on the documentation.
We will generate a one-dimensional dataset with a simple model (including some noise), and we will try to fit a function to this data. With this function, we can predict values on new data points. This is a curve-fitting regression problem.
import numpy as np
import scipy.stats as st
import sklearn.linear_model as lm
import matplotlib.pyplot as plt
%matplotlib inline
f = lambda x: np.exp(3 * x)
x_tr = np.linspace(0., 2, 200)
y_tr = f(x_tr)
x = np.array([0, .1, .2, .5, .8, .9, 1])
y = f(x) + np.random.randn(len(x))
plt.figure(figsize=(6,3));
plt.plot(x_tr[:100], y_tr[:100], '--k');
plt.plot(x, y, 'ok', ms=10);
LinearRegression
class). Then we fit the model to our data. Finally, we predict values from our trained model.# We create the model.
lr = lm.LinearRegression()
# We train the model on our training dataset.
lr.fit(x[:, np.newaxis], y);
# Now, we predict points with our trained model.
y_lr = lr.predict(x_tr[:, np.newaxis])
We need to convert x
and x_tr
to column vectors, as it is a general convention in scikit-learn that observations are rows, while features are columns. Here, we have 7 observations with 1 feature.
plt.figure(figsize=(6,3));
plt.plot(x_tr, y_tr, '--k');
plt.plot(x_tr, y_lr, 'g');
plt.plot(x, y, 'ok', ms=10);
plt.xlim(0, 1);
plt.ylim(y.min()-1, y.max()+1);
plt.title("Linear regression");
np.vander
function. We will explain this trick in more detail in How it works....lrp = lm.LinearRegression()
plt.figure(figsize=(6,3));
plt.plot(x_tr, y_tr, '--k');
for deg, s in zip([2, 5], ['-', '.']):
lrp.fit(np.vander(x, deg + 1), y);
y_lrp = lrp.predict(np.vander(x_tr, deg + 1))
plt.plot(x_tr, y_lrp, s, label='degree ' + str(deg));
plt.legend(loc=2);
plt.xlim(0, 1.4);
plt.ylim(-10, 40);
# Print the model's coefficients.
print(' '.join(['%.2f' % c for c in lrp.coef_]))
plt.plot(x, y, 'ok', ms=10);
plt.title("Linear regression");
22.03 -4.25 0.00 -476.66 1286.40 -1171.56 419.94 -38.56 0.00
We have fitted two polynomial models of degree 2 and 5. The degree 2 polynomial appears to fit the data points less precisely than the degree 5 polynomial. However, it seems more robust: the degree 5 polynomial seems really bad at predicting values outside the data points (look for example at the portion $x >= 1$). This is what we call overfitting: by using a too complex model, we obtain a better fit on the trained dataset, but a less robust model outside this set. Note the large coefficients of the degree 5 polynomial: this is generally a sign of overfitting.
The ridge regression model has a meta-parameter which represents the weight of the regularization term. We could try different values with trials and errors, using the Ridge
class. However, scikit-learn includes another model called RidgeCV
which includes a parameter search with cross-validation. In practice, it means that we don't have to tweak this parameter by hand: scikit-learn does it for us. Since the models of scikit-learn always follow the fit
-predict
API, all we have to do is replace lm.LinearRegression
by lm.RidgeCV
in the code above. We will give more details in the next section.
ridge = lm.RidgeCV()
plt.figure(figsize=(6,3));
plt.plot(x_tr, y_tr, '--k');
for deg, s in zip([2, 5], ['-', '.']):
ridge.fit(np.vander(x, deg + 1), y);
y_ridge = ridge.predict(np.vander(x_tr, deg + 1))
plt.plot(x_tr, y_ridge, s, label='degree ' + str(deg));
plt.legend(loc=2);
plt.xlim(0, 1.5);
plt.ylim(-5, 80);
# Print the model's coefficients.
print(' '.join(['%.2f' % c for c in ridge.coef_]))
plt.plot(x, y, 'ok', ms=10);
plt.title("Ridge regression");
11.26 5.93 0.00 4.39 3.80 3.32 3.27 3.97 0.00
This time, the degree 5 polynomial seems more precise than the simpler degree 2 polynomial (which now causes underfitting). Ridge regression reduces the overfitting issue here. Observe how the degree 5 polynomial's coefficients are much smaller than in the previous example.
In this section, we will explain all the aspects covered in the example above.
scikit-learn implements a clean and coherent API for supervised and unsupervised learning. Our data points should be in a $N \times D$ matrix $\mathbf{X}$, where $N$ is the number of observations, and $D$ is the number of features. In other words, each row is an observation. The first step in a machine learning task is to define what our matrix $\mathbf{X}$ is exactly.
In a supervised learning setup, we also have a target, a $N$-long vector $\mathbf{y}$ with a scalar value for each observations. This value is continuous or discrete, depending on whether we have a regression or classification problem, respectively.
In scikit-learn, models are implemented in classes that have fit
and predict
methods. The fit
method accepts the data matrix X
as input, and y
as well for supervised learning models. This method trains the model on the given data.
The predict
method also takes data points as input (as a matrix $M \times D$). It returns the labels or transformed points, according to the trained model.
Ordinary least squares regression is one of the simplest regression methods. It consists of modeling the output values $\hat{y}_i$ as a linear combination of $X_{ij}$:
$$\forall i \in \{1, \ldots, N\}, \quad \hat{y}_i = \sum_{j=1}^D w_j X_{ij}, \quad \textrm{or, in matrix form:} \quad \mathbf{\hat{y}} = \mathbf{X} \mathbf{w}.$$Here, $\mathbf{w} = (w_1, \ldots, w_D)$ is the (unknown) parameter vector. Also, $\mathbf{\hat{y}}$ represents the model's output. We want this vector to match the data points $\mathbf{y}$ as closely as possible. Of course, the exact equality $\mathbf{\hat{y}} = \mathbf{y}$ cannot hold in general (there is always some amount of noise in real-world data). Therefore, we want to minimize the difference between those two vectors. The ordinary least squares regression method consists of minimizing the following loss function:
$$\min_{\mathbf{w}} || \mathbf{y} - \mathbf{X} \mathbf{w} ||_2^2 = \min_{\mathbf{w}} \left( \sum_{i=1}^N \left(y_i - \hat{y}_i\right)^2 \right)$$This sum of the components squared is called the $l_2$ norm (or Euclidean norm). It is convenient because it leads to differentiable loss functions, so that gradients can be computed and common optimization procedures can be performed.
Ordinary Least Squares regression fits a linear model to the data. The model is linear both in the data points $X_i$ and in the parameters $w_j$. In our example, we obtained a poor fit because the data points were generated according to a nonlinear generative model (an exponential function).
However, we can still use the linear regression method with a model that is linear in $w_j$, but nonlinear in $\mathbf{x}_i$. To do this, we need to increase the number of dimensions in our dataset by using a basis of polynomial functions. In other words, we consider the following observations:
$$\mathbf{x}_i, \mathbf{x}_i^2, \ldots, \mathbf{x}_i^D$$where $D$ is the maximum degree. The input matrix $\mathbf{X}$ is therefore the Vandermonde matrix associated to the original data points $\mathbf{x}_i$.
Polynomial interpolation with linear regression can lead to overfitting issues if the degree of the polynomials is too large. By capturing the random fluctuations (noise) instead of the general trend of the data, the model's predictive power decreases. This corresponds to an explosion of the polynomial's coefficients $w_j$.
A solution to this problem is to prevent these coefficients from growing unboundedly. With ridge regression, this is done by adding a regularization term to the loss function:
$$\min_{\mathbf{w}} || \mathbf{y} - \mathbf{X} \mathbf{w} ||_2^2 + \alpha ||\mathbf{w}||_2^2$$By minimizing this loss function, we not only minimize the error between the model and the data (first term, related to the bias), but also the size of the model's coefficients (second term, related to the variance). The bias-variance trade-off is quantified by the hyperparameter $\alpha$, which precises the relative weight between the two terms in the loss function.
Here, ridge regression led to a polynomial with smaller coefficients, and thus a better fit.
A drawback of the ridge regression model compared to the ordinary least squares model is the presence of an extra hyperparameter $\alpha$. The quality of the prediction depends on the choice of this parameter. One possibility would be to fine-tune this parameter manually, but this procedure can be tedious and can also lead to overfitting.
We can use a grid search for this: we loop over many possible values for $\alpha$, and we evaluate the performance of the model for each possible value. Then, we choose the parameter that yields the best performance.
How to assess the performance of a model with a given $\alpha$ value? A common solution is to use cross-validation. This procedure consists of splitting the dataset into a train set and a test set. We fit the model on the train set, and test its predictive performance on the test set. By testing the model on a different dataset than the one used for training, we avoid overfitting.
There are many ways to split the initial dataset into two parts like this. One possibility is to remove one sample to form the train set, and to put this one sample into the test set. This is called Leave-One-Out cross-validation. With $N$ samples, we obtain $N$ sets of train and test sets. The cross-validated performance is the average performance on all these set decompositions.
As we will see in Chapter 8, scikit-learn implements several easy-to-use functions to do cross-validation and grid search. Here, we used a special estimator called RidgeCV
that implements a cross-validation and grid search procedure that is specific to the ridge regression model. Using this model ensures that the best hyperparameter $\alpha$ is found automatically for us.
In the book, the rest of the chapter introduces many standard algorithms in machine learning.
Here are a few references about cross-validation and grid search:
Here are a few references about scikit-learn:
Here are a few excellent, math-heavy textbooks on machine learning:
Here are a few books targetting programmers without necessarily a strong mathematical background:
You will find many other references online.
This was a featured recipe from the IPython Cookbook, by Cyrille Rossant, Packt Publishing, 2014.