Kevin Markham
Uploaded Nov 25, 2015

Beginner's Guide to Click-Through Rate Prediction with Logistic Regression

Let's say that you're a major search engine, and you need to decide which ad to display at the top of your search results. How would you do it?

Your first thought might be to narrow the scope to ads "related" to the search, and then choose whichever ad offers the greatest revenue. Companies have already bid on how much they will pay you, so it seems easy to maximize your revenue by choosing the highest paying ad. But is that the right approach?

Many ads are actually sold on a "pay-per-click" (PPC) basis, meaning the company only pays for ad clicks, not ad views. Thus your optimal approach (as a search engine) is actually to choose an ad based on "expected value", meaning the price of a click times the likelihood that the ad will be clicked. In other words, a \$1.00 ad with a 5% probability of being clicked has an expected value of \$0.05, whereas a \$2.00 ad with a 1% probability of being clicked has an expected value of only \$0.02. In this case, you would choose to display the first ad.

In order for you to maximize expected value, you therefore need to accurately predict the likelihood that a given ad will be clicked, also known as "click-through rate" (CTR).

In this notebook, I'll walk through the predictive modeling process, discuss why logistic regression is a good choice for this task, and then explain this code line-by-line so that you can apply it to your own predictive task!

For this example, I'm using the data from a Kaggle competition on click-through rate prediction sponsored by Avazu. The goal in the competition matches our goal, which is to predict the likelihood that a given ad will be clicked.

Step 1: Reading and Exploring the Data

I've already downloaded the dataset from Kaggle for this example and extracted a small subset to make my calculations faster. If you would like to follow along, you should download and decompress train.gz from the competition's data page (login required), and then extract the first 100,000 lines from train.csv using this command at the command line/terminal: head -n100000 train.csv > train_subset.csv

Our first step is to read the data into an SFrame, which is GraphLab's tabular data structure that is similar to a data frame in R or a pandas DataFrame in Python.

This data happens to be stored in the popular CSV (comma separated value) format, but SFrames can be constructed from a variety of sources. We'll use the read_csv method to read in the data:

In [1]:
import graphlab as gl
data = gl.SFrame.read_csv('train_subset.csv', verbose=False)

Let's take a quick look at the first row of data, to see what we're working with:

In [2]:
id click hour C1 banner_pos site_id site_domain site_category app_id
1000009418151094273 0 14102100 1005 0 1fbe01fe f3845767 28905ebd ecad2386
app_domain app_category device_id device_ip device_model device_type device_conn_type C14 C15 C16
7801e8d9 07d7df22 a99f214a ddd2926e 44956a24 1 2 15706 320 50
C17 C18 C19 C20 C21
1722 0 35 -1 79
[1 rows x 24 columns]

From Kaggle's data dictionary, I know that click=0 means the ad was not clicked, and click=1 means the ad was clicked. The "click" column is therefore our target variable, and the other columns are our potential features!

The first thing we want to know is what percentage of ads in the dataset were actually clicked. In this case, we can simply take the mean of the "click" column, since that is equivalent to adding up all of the ones (which is the number of clicks) and dividing by the total number of ads:

In [3]:

We see that 17.5% of the ads were clicked, meaning the overall click-through rate is 17.5%. This is useful to keep in mind as a "baseline", as we'll see later on.

Before we start building a machine learning model, it's always useful to explore the dataset. One way to get started is by using the GraphLab Canvas, a browser-based visualization platform:

In [4]:

I noticed that "device_type" only has 4 unique values, and it makes intuitive sense that the type of device you're using when viewing an ad might affect your likelihood of clicking the ad, so let's explore it further.

To understand the relationship between this feature and the target variable, we want to calculate the click-through rate for each value of device_type. We can accomplish this by "grouping the data" by device_type, and then calculating the mean of the click column for each group:

In [5]:
data.groupby('device_type', {'CTR':gl.aggregate.MEAN('click')})
device_type CTR
0 0.227499406317
5 0.0990566037736
4 0.0725075528701
1 0.175977623465
[4 rows x 2 columns]

We saw earlier that the baseline click-through rate is 17.5%, and it appears that there is a big difference in average click-through rate depending on device_type. This looks like a good feature!

Similarly, the C1 column looks like a good feature:

In [6]:
data.groupby('C1', {'CTR':gl.aggregate.MEAN('click')})
1008 0.4
1005 0.176174097389
1001 0.103448275862
1010 0.0742713882795
1002 0.227499406317
1007 0.0
[6 rows x 2 columns]

I also noticed that C15 and C16 appear to be the dimensions of the ad (width and height), which we would also imagine are good predictors of whether an ad is clicked:

In [7]:
{120: 2, 216: 912, 300: 3935, 320: 95132, 728: 18}
In [8]:
{20: 2, 36: 912, 50: 95620, 90: 18, 250: 3427, 480: 20}

For our initial model, we'll just use device_type, C1, C15, and C16 as our features.

Note that when we built the SFrame from the CSV file, it simply guessed the data type of each column. Sometimes these data types need to be adjusted, so let's take a quick look at the column names and their associated types to see if there's anything we need to fix:

In [9]:
zip(data.column_names(), data.column_types())
[('id', str),
 ('click', int),
 ('hour', int),
 ('C1', int),
 ('banner_pos', int),
 ('site_id', str),
 ('site_domain', str),
 ('site_category', str),
 ('app_id', str),
 ('app_domain', str),
 ('app_category', str),
 ('device_id', str),
 ('device_ip', str),
 ('device_model', str),
 ('device_type', int),
 ('device_conn_type', int),
 ('C14', int),
 ('C15', int),
 ('C16', int),
 ('C17', int),
 ('C18', int),
 ('C19', int),
 ('C20', int),
 ('C21', int)]

We know that both device_type and C1 are "categorical variables", meaning that their numerical values represent categories. We'll convert the data type of both of those columns from integer to string, because we don't want our machine learning model to think there is a mathematical relationship between the category values:

In [10]:
data['device_type'] = data['C1'].astype(str)
data['C1'] = data['C1'].astype(str)

You could spend a lot more time on the exploratory phase, but let's move along to the next step in predictive modeling! If you want to learn how to manipulate SFrames in more depth, read through this example notebook, Introduction to SFrames.

Step 2: Splitting the Data

One of the keys to proper machine learning is model evaluation. The goal of model evaluation is to estimate how well your model will "generalize" to future data. In other words, we want to build a model that accurately predicts the future, not the past!

One of the most common evaluation procedures is to split your data into a "training set" and a "testing set".

Let's use an 80/20 split, in which 80% of the data is used for training and 20% is used for testing:

In [11]:
train_data, test_data = data.random_split(0.8, seed=1)

We now have two separate SFrames, called train_data and test_data.

Step 3: Selecting a Machine Learning Model

There are two main types of models: classification models, which are used when your target variable is categorical (such as yes/no), and regression models, which are used when your target variable is continuous (such as price). In this case, we'll need to use a classification model, since our target variable is categorical (click: yes or no).

The specific model we're going to use in this case is logistic regression. In logistic regression, the probability that the target is True is modeled as a logistic function of a linear combination of the features. Thus, the model is predicting a probability (which is a continuous value), but that probability is used to choose the predicted target class. In other words, it's using regression to predict a continuous value, but we're using the continuous value that is output from the model to perform classification. (Pretty cool, right?)

It can take a lot of study to truly understand a machine learning model, but a good introduction to logistic regression is available in the user guide.

So, why exactly did we choose logistic regression for this task, instead of any of the other available classification models? Well, it turns out that logistic regression has many nice properties. For starters, it is a very fast model, meaning that it does not take long to train the model or make predictions. As well, it is highly interpretable, meaning that you can understand exactly how it's making predictions. But the key consideration in this case is that logistic regression outputs "well-calibrated" predicted probabilities.

Step 4: Training a Machine Learning Model

Now that we've selected our model, we can finally start the model training process! In GraphLab Create, this can be done in a single line. You simply pass in the training data, the name of the target column, and the names of the feature columns. And in fact, if you just replace gl.logistic_classifier.create with gl.classifier.create, GraphLab will choose the best model for you automatically (based on the properties of your data), meaning that you can skip Step 3 above!

In [12]:
model = gl.logistic_classifier.create(train_data, target='click', features=['device_type', 'C1', 'C15', 'C16'])
PROGRESS: Creating a validation set from 5 percent of training data. This may take a while.
          You can set ``validation_set=None`` to disable validation tracking.

PROGRESS: Logistic regression:
PROGRESS: --------------------------------------------------------
PROGRESS: Number of examples          : 76149
PROGRESS: Number of classes           : 2
PROGRESS: Number of feature columns   : 4
PROGRESS: Number of unpacked features : 4
PROGRESS: Number of coefficients    : 13
PROGRESS: Starting Newton Method
PROGRESS: --------------------------------------------------------
PROGRESS: +-----------+----------+--------------+-------------------+---------------------+
PROGRESS: | Iteration | Passes   | Elapsed Time | Training-accuracy | Validation-accuracy |
PROGRESS: +-----------+----------+--------------+-------------------+---------------------+
PROGRESS: | 1         | 2        | 1.074658     | 0.824095          | 0.819668            |
PROGRESS: | 2         | 3        | 1.122571     | 0.824095          | 0.819668            |
PROGRESS: | 3         | 4        | 1.168321     | 0.824095          | 0.819668            |
PROGRESS: | 4         | 5        | 1.224780     | 0.824095          | 0.819668            |
PROGRESS: | 5         | 6        | 1.274098     | 0.824095          | 0.819668            |
PROGRESS: | 6         | 7        | 1.319348     | 0.824095          | 0.819668            |
PROGRESS: +-----------+----------+--------------+-------------------+---------------------+

PROGRESS: TERMINATED: Iteration limit reached.
PROGRESS: This model may not be optimal. To improve it, consider increasing `max_iterations`.

Note that we didn't have to tell GraphLab how to handle each of the features, even though two features were numerical and the other two were categorical. The categorical features were automatically handled using "dummy encoding", which is why the output above indicates that there were 4 features but 13 model coefficients. (A simple explanation of dummy encoding is available in the user guide.)

Step 5: Making Predictions

After training a model, the final step is to use the model to make predictions. In other words, the model has learned a mathematical relationship between the features and the target, and it will use that relationship to predict the target value for new data points.

In this case, we pass the testing data to the "fitted model", and ask it to output the predicted probability of a click:

In [13]:
model.predict(test_data, output_type='probability').head(5)
dtype: float
Rows: 5
[0.16537085227336723, 0.22480874210027335, 0.16537085227336723, 0.16537085227336723, 0.16537085227336723]

At this point, you would want to evaluate the model by comparing the predicted probabilities versus the actual target values, using an appropriate "evaluation metric." The best metric to use in this case is probably logarithmic loss, which is commonly used when you care about having well-calibrated probabilities. In addition, you might inspect the ROC curve and compute other metrics such as the F1-score and AUC. (See this blog post for more details.)

Now that we have these probabilities, we could find the ad that maximizes revenue by multiplying these probabilities by the cost-per-click, and finding the largest value.

Although we're at the end of this notebook, this is really just the beginning! You can continue to add more features to the model, and then use the evaluation metric to compare the expected performance of each of your models. As well, you can use feature engineering to create new features, you can try other models, and so much more!

If you'd like to read more about click-through rate prediction, there are readable papers on the topic by both Criteo and Google.