Author: Yury Kashnitskiy. Translated and edited by Christina Butsko, Nerses Bagiyan, Yulia Klimushina, and Yuanyuan Pao. This material is subject to the terms and conditions of the Creative Commons CC BY-NC-SA 4.0 license. Free use is permitted for any non-commercial purpose.

Now for a little practice! We want to solve the problem of binary classification of IMDB movie reviews. We have a training set with marked reviews, 12500 reviews marked as good, another 12500 bad. Here, it's not easy to get started with machine learning right away because we don't have the matrix $X$; we need to prepare it. We will use a simple approach: bag of words model. Features of the review will be represented by indicators of the presence of each word from the whole corpus in this review. The corpus is the set of all user reviews. The idea is illustrated by a picture

In [1]:

```
import os
import numpy as np
import matplotlib.pyplot as plt
import numpy as np
from sklearn.datasets import load_files
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.linear_model import LogisticRegression
```

In [2]:

```
# Uncomment this to load data from downloaded tar.gz archive.
# change this path
PATH_TO_IMDB = '/Users/y.kashnitsky/Documents/Machine_learning/datasets/aclImdb/'
reviews_train = load_files(os.path.join(PATH_TO_IMDB, "train"),
categories=['pos', 'neg'])
text_train, y_train = reviews_train.data, reviews_train.target
# change the path to the file
reviews_test = load_files(os.path.join(PATH_TO_IMDB, "test"),
categories=['pos', 'neg'])
text_test, y_test = reviews_test.data, reviews_test.target
```

In [3]:

```
# # Alternatively, load data from previously pickled objects.
# import pickle
# with open('../../data/imdb_text_train.pkl', 'rb') as f:
# text_train = pickle.load(f)
# with open('../../data/imdb_text_test.pkl', 'rb') as f:
# text_test = pickle.load(f)
# with open('../../data/imdb_target_train.pkl', 'rb') as f:
# y_train = pickle.load(f)
# with open('../../data/imdb_target_test.pkl', 'rb') as f:
# y_test = pickle.load(f)
```

In [4]:

```
print("Number of documents in training data: %d" % len(text_train))
print(np.bincount(y_train))
print("Number of documents in test data: %d" % len(text_test))
print(np.bincount(y_test))
```

**Here are a few examples of the reviews.**

In [5]:

```
print(text_train[1])
```

In [6]:

```
y_train[1] # bad review
```

Out[6]:

In [7]:

```
text_train[2]
```

Out[7]:

In [8]:

```
y_train[2] # good review
```

Out[8]:

In [9]:

```
# import pickle
# with open('../../data/imdb_text_train.pkl', 'wb') as f:
# pickle.dump(text_train, f)
# with open('../../data/imdb_text_test.pkl', 'wb') as f:
# pickle.dump(text_test, f)
# with open('../../data/imdb_target_train.pkl', 'wb') as f:
# pickle.dump(y_train, f)
# with open('../../data/imdb_target_test.pkl', 'wb') as f:
# pickle.dump(y_test, f)
```

**First, we will create a dictionary of all the words using CountVectorizer**

In [10]:

```
cv = CountVectorizer()
cv.fit(text_train)
len(cv.vocabulary_)
```

Out[10]:

**If you look at the examples of "words" (let's call them tokens), you can see that we have omitted many of the important steps in text processing (automatic text processing can itself be a completely separate series of articles).**

In [11]:

```
print(cv.get_feature_names()[:50])
print(cv.get_feature_names()[50000:50050])
```

**Secondly, we are encoding the sentences from the training set texts with the indices of incoming words. We'll use the sparse format.**

In [12]:

```
X_train = cv.transform(text_train)
X_train
```

Out[12]:

**Let's see how our transformation worked**

In [13]:

```
print(text_train[19726])
```

In [14]:

```
X_train[19726].nonzero()[1]
```

Out[14]:

In [15]:

```
X_train[19726].nonzero()
```

Out[15]:

**Third, we will apply the same operations to the test set**

In [16]:

```
X_test = cv.transform(text_test)
```

**The next step is to train Logistic Regression.**

In [17]:

```
%%time
logit = LogisticRegression(solver='lbfgs', n_jobs=-1, random_state=7)
logit.fit(X_train, y_train)
```

**Let's look at accuracy on the both the training and the test sets.**

In [18]:

```
round(logit.score(X_train, y_train), 3), round(logit.score(X_test, y_test), 3),
```

Out[18]:

**The coefficients of the model can be beautifully displayed.**

In [19]:

```
def visualize_coefficients(classifier, feature_names, n_top_features=25):
# get coefficients with large absolute values
coef = classifier.coef_.ravel()
positive_coefficients = np.argsort(coef)[-n_top_features:]
negative_coefficients = np.argsort(coef)[:n_top_features]
interesting_coefficients = np.hstack([negative_coefficients, positive_coefficients])
# plot them
plt.figure(figsize=(15, 5))
colors = ["red" if c < 0 else "blue" for c in coef[interesting_coefficients]]
plt.bar(np.arange(2 * n_top_features), coef[interesting_coefficients], color=colors)
feature_names = np.array(feature_names)
plt.xticks(np.arange(1, 1 + 2 * n_top_features), feature_names[interesting_coefficients], rotation=60, ha="right");
```

In [20]:

```
def plot_grid_scores(grid, param_name):
plt.plot(grid.param_grid[param_name], grid.cv_results_['mean_train_score'],
color='green', label='train')
plt.plot(grid.param_grid[param_name], grid.cv_results_['mean_test_score'],
color='red', label='test')
plt.legend();
```

In [21]:

```
visualize_coefficients(logit, cv.get_feature_names())
```

**To make our model better, we can optimize the regularization coefficient for the Logistic Regression. We'll use sklearn.pipeline because CountVectorizer should only be applied to the training data (so as to not "peek" into the test set and not count word frequencies there). In this case, pipeline determines the correct sequence of actions: apply CountVectorizer, then train Logistic Regression.**

In [27]:

```
%%time
from sklearn.pipeline import make_pipeline
text_pipe_logit = make_pipeline(CountVectorizer(),
# for some reason n_jobs > 1 won't work
# with GridSearchCV's n_jobs > 1
LogisticRegression(solver='lbfgs',
n_jobs=1,
random_state=7))
text_pipe_logit.fit(text_train, y_train)
print(text_pipe_logit.score(text_test, y_test))
```

In [28]:

```
%%time
from sklearn.model_selection import GridSearchCV
param_grid_logit = {'logisticregression__C': np.logspace(-5, 0, 6)}
grid_logit = GridSearchCV(text_pipe_logit, param_grid_logit,
cv=3, n_jobs=-1)
grid_logit.fit(text_train, y_train)
```

**Let's print best $C$ and cv-score using this hyperparameter:**

In [29]:

```
grid_logit.best_params_, grid_logit.best_score_
```

Out[29]:

In [30]:

```
plot_grid_scores(grid_logit, 'logisticregression__C')
```

For the validation set:

In [31]:

```
grid_logit.score(text_test, y_test)
```

Out[31]:

**Now let's do the same with random forest. We see that, with logistic regression, we achieve better accuracy with less effort.**

In [32]:

```
from sklearn.ensemble import RandomForestClassifier
```

In [33]:

```
forest = RandomForestClassifier(n_estimators=200,
n_jobs=-1, random_state=17)
```

In [34]:

```
%%time
forest.fit(X_train, y_train)
```

Out[34]:

In [35]:

```
round(forest.score(X_test, y_test), 3)
```

Out[35]:

Let's now consider an example where linear models are worse.

Linear classification methods still define a very simple separating surface - a hyperplane. The most famous toy example of where classes cannot be divided by a hyperplane (or line) with no errors is "the XOR problem".

XOR is the "exclusive OR", a Boolean function with the following truth table:

XOR is the name given to a simple binary classification problem in which the classes are presented as diagonally extended intersecting point clouds.

In [36]:

```
# creating dataset
rng = np.random.RandomState(0)
X = rng.randn(200, 2)
y = np.logical_xor(X[:, 0] > 0, X[:, 1] > 0)
```

In [37]:

```
plt.scatter(X[:, 0], X[:, 1], s=30, c=y, cmap=plt.cm.Paired);
```

Obviously, one cannot draw a single straight line to separate one class from another without errors. Therefore, logistic regression performs poorly with this task.

In [38]:

```
def plot_boundary(clf, X, y, plot_title):
xx, yy = np.meshgrid(np.linspace(-3, 3, 50),
np.linspace(-3, 3, 50))
clf.fit(X, y)
# plot the decision function for each datapoint on the grid
Z = clf.predict_proba(np.vstack((xx.ravel(), yy.ravel())).T)[:, 1]
Z = Z.reshape(xx.shape)
image = plt.imshow(Z, interpolation='nearest',
extent=(xx.min(), xx.max(), yy.min(), yy.max()),
aspect='auto', origin='lower', cmap=plt.cm.PuOr_r)
contours = plt.contour(xx, yy, Z, levels=[0], linewidths=2,
linetypes='--')
plt.scatter(X[:, 0], X[:, 1], s=30, c=y, cmap=plt.cm.Paired)
plt.xticks(())
plt.yticks(())
plt.xlabel(r'$x_1$')
plt.ylabel(r'$x_2$')
plt.axis([-3, 3, -3, 3])
plt.colorbar(image)
plt.title(plot_title, fontsize=12);
```

In [39]:

```
plot_boundary(LogisticRegression(), X, y,
"Logistic Regression, XOR problem")
```

But if one were to give polynomial features as an input (here, up to 2 degrees), then the problem is solved.

In [40]:

```
from sklearn.preprocessing import PolynomialFeatures
from sklearn.pipeline import Pipeline
```

In [41]:

```
logit_pipe = Pipeline([('poly', PolynomialFeatures(degree=2)),
('logit', LogisticRegression())])
```

In [42]:

```
plot_boundary(logit_pipe, X, y,
"Logistic Regression + quadratic features. XOR problem")
```

Here, logistic regression has still produced a hyperplane but in a 6-dimensional feature space $1, x_1, x_2, x_1^2, x_1x_2$ and $x_2^2$. When we project to the original feature space, $x_1, x_2$, the boundary is nonlinear.

In practice, polynomial features do help, but it is computationally inefficient to build them explicitly. SVM with the kernel trick works much faster. In this approach, only the distance between the objects (defined by the kernel function) in a high dimensional space is computed, and there is no need to produce a combinatorially large number of features.

- Main course site, course repo, and YouTube channel
- Medium "story" based on this notebook
- Course materials as a Kaggle Dataset
- If you read Russian: an article on Habrahabr with ~ the same material. And a lecture on YouTube
- A nice and concise overview of linear models is given in the book “Deep Learning” (I. Goodfellow, Y. Bengio, and A. Courville).
- Linear models are covered practically in every ML book. We recommend “Pattern Recognition and Machine Learning” (C. Bishop) and “Machine Learning: A Probabilistic Perspective” (K. Murphy).
- If you prefer a thorough overview of linear model from a statistician’s viewpoint, then look at “The elements of statistical learning” (T. Hastie, R. Tibshirani, and J. Friedman).
- The book “Machine Learning in Action” (P. Harrington) will walk you through implementations of classic ML algorithms in pure Python.
- Scikit-learn library. These guys work hard on writing really clear documentation.
- Scipy 2017 scikit-learn tutorial by Alex Gramfort and Andreas Mueller.
- One more ML course with very good materials.
- Implementations of many ML algorithms. Search for linear regression and logistic regression.