# Logistic Regression

Logistic regression is a regression model that is popularly used for classification tasks. In logistic regression, the probability that a binary target is True is modeled as a logistic function of a linear combination of features.

The following figure illustrates how logistic regression is used to create a 1-dimensional classifier. The training data consists of positive examples (depicted in blue) and negative examples (in orange). The decision boundary (depicted in pink) separates out the data into two classes.

##### Background

Given a set of features , and a label , logistic regression interprets the probability that the label is in one class as a logistic function of a linear combination of the features:

Analogous to linear regression, an intercept term is added by appending a column of 1's to the features and L1 and L2 regularizers are supported. The composite objective being optimized for is the following:

where is the L1_penalty and is the L2_penalty.

##### Introductory Example

First, let's construct a binary target variable. In this example, we will predict if a restaurant is good or bad, with 1 and 2 star ratings indicating a bad business and 3-5 star ratings indicating a good one. We will use the following features.

• Average rating of a given business
• Average rating made by a user
• Number of reviews made by a user
• Number of reviews that concern a business
import turicreate as tc

data =  tc.SFrame('ratings-data.csv')

# Restaurants with rating >=3 are good
data['is_good'] = data['stars'] >= 3

# Make a train-test split
train_data, test_data = data.random_split(0.8)

# Create a model.
model = tc.logistic_classifier.create(train_data, target='is_good',
features = ['user_avg_stars',
'user_review_count',

# Save predictions (probability estimates) to an SArray
predictions = model.classify(test_data)

# Evaluate the model and save the results into a dictionary
results = model.evaluate(test_data)

Refer to the chapter on linear regression for the following features:

We will now discuss some advanced features that are specific to logistic regression.

###### Making Predictions

Predictions using a Turi Create classifier is easy. The classify() method provides a one-stop shop for all that you need from a classifier.

• A class prediction
• Probability/Confidence associated with that class prediction.

In the following example, the first prediction was class 1 with a 90.5% probability.

predictions = model.classify(test_data)
print(predictions)
+-------+----------------+
| class |  probability   |
+-------+----------------+
|   1   | 0.905618772131 |
|   1   | 0.941576249302 |
|   1   | 0.948254387657 |
|   0   | 0.996952633956 |
|   1   | 0.944229260472 |
|   1   | 0.951769966846 |
|   1   | 0.905561314917 |
|   1   | 0.957697429248 |
|   1   | 0.98527411871  |
|   1   | 0.973282185166 |
|  ...  |      ...       |
+-------+----------------+
[43018 rows x 2 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.
###### Making detailed predictions

Logistic regression predictions can take one of three forms:

• Classes (default): Thresholds the probability estimate at 0.5 to predict
• a class label i.e. 0/1.
• Probabilities: A probability estimate (in the range [0,1]) that the example is in the True class. Note that this is not the same as the probability estimate in the classify function.
• Margins : Distance to the linear decision boundary learned by the model. The larger the distance, the more confidence we have that it belongs to one class or the other.

Turi Create's logistic regression model can return predictions for any of these types:

pred_class = model.predict(test_data, output_type = "class")          # Class
pred_prob_one = model.predict(test_data, output_type = 'probability') # Probability
pred_margin = model.predict(test_data, output_type = "margin")        # Margins
###### Evaluating Results

We can also evaluate our predictions by comparing them to known ratings. The results are evaluated using two metrics:

result = model.evaluate(test_data)
print(Accuracy         : %s" % result['accuracy'])
print("Confusion Matrix : \n%s" % result['confusion_matrix'])
Accuracy         : 0.860862092991
Confusion Matrix :
+--------------+-----------------+-------+
| target_label | predicted_label | count |
+--------------+-----------------+-------+
|      0       |        0        |  2348 |
|      0       |        1        |  4816 |
|      1       |        0        |  912  |
|      1       |        1        | 34942 |
+--------------+-----------------+-------+
[4 rows x 3 columns]

Using basic SFrame operations, we can also isolate the examples in the test data where the model made mistakes:

predictions = model.predict(test_data)

# Compute a boolean SArray of whether or not the model was right
mistakes_filter = predictions != test_data[model.target]
correct_filter = predictions == test_data[model.target]

# Apply the logical filter on the data
mistakes = test_data[mistakes_filter]
correct = test_data[correct_filter]

We can use the similar idea to isolate mistakes that are:

• false positives: The model predicted true (for a class) but the ground truth was false.
• false negative: The model predicted false (for a class) but the ground truth was true.
predictions = model.predict(test_data)

# Compute boolean filters
false_positive_filter = (predictions == 1) & (test_data[model.target] == 0)
false_negative_filter = (predictions == 0) & (test_data[model.target] == 1)

false_negatives = test_data[false_negative_filter]
false_positives = test_data[false_positive_filter]
###### Working with imbalanced data

Many difficult real-world problems have imbalanced data, where at least one class is under-represented. Turi Create models can handle the imbalanced data by assigning asymmetric costs of misclassifying elements of different classes.

# Load the data

# Label 'c' is edible
data['label'] = data['label'] == 'c'

# Make a train-test split
train_data, test_data = data.random_split(0.8)

# Create a model which weights classes based on frequency in the training data.
model = tc.logistic_classifier.create(train_data, target='label',
class_weights = 'auto')
##### Multiclass Classification

Multiclass classification is the problem of classifying instances into one of many (i.e more than two) possible instances. As an example, binary classification can be used to create a classifier that can distinguish between two classes, say "cat" or "dog", while multiclass classification can be used to create a model a finite set of labels at the same time, say "cat", "dog", "rat", and "cow".

import turicreate as tc

data = tc.SFrame('mnist/train6k-array')

# Make a train-test split
train_data, test_data = data.random_split(0.8)

# Create a model
model = tc.logistic_classifier.create(train_data, target='label')

# Save predictions to an SFrame (class and corresponding class-probabilities)
predictions = model.classify(test_data)

# Top 5 predictions with probabilities, rank, and margin
top = model.predict_topk(test_data, output_type='probability', k = 5)
top = model.predict_topk(test_data, output_type='rank', k = 5)
top = model.predict_topk(test_data, output_type='margin', k = 5)

# Evaluate the model and save the results into a dictionary
results = model.evaluate(test_data)
###### Top-k predictions

Multiclass classification provides the top-k class predictions for each class. The predictions are either margins, probabilities, or a rank for the predicted class for each example. In the following example, we provide the top 5 predictions, ordered by class probability, for each data point in the test set.

top = model.predict_topk(test_data, output_type='probability', k = 3)
print(top)
Columns:
id  str
class str
probability float

Rows: 3711

Data:
+-----+-------+------------------+
|  id | class |   probability    |
+-----+-------+------------------+
|  0  |   5   |  0.94296406887   |
|  0  |   1   | 0.0140330526641  |
|  0  |   8   | 0.00636249767982 |
|  1  |   8   |  0.929146865934  |
|  1  |   9   | 0.0139581314344  |
|  1  |   1   | 0.00982837828507 |
|  2  |   1   |  0.937192457289  |
|  2  |   7   | 0.0106293228679  |
|  2  |   4   | 0.00910849289074 |
|  3  |   5   |  0.900146607924  |
| ... |  ...  |       ...        |
+-----+-------+------------------+