Ad Click Prediction

This repository contains the Python scripts needed to predict whether a user will click on an ad.

Data

We use the well known iPinYou dataset to train our classifier. The dataset was created during a Real-Time Bidding competition. We use a modified version of this dataset, which contains data for 4,000,000 individual ad bids. For our classifications, we use the following variables:

Name	Type
click	boolean
imp	boolean
Browser	categorical
AdExchange	categorical
Adslotvisibility	categorical
Adslotformat	categorical
Payingprice	int
Adslotwidth	int
Adslotheight	int
interest_*	boolean
Inmarket_*	boolean
Demographic_gender_female	boolean

Obviously, click is the label variable, which our classifier will regress. imp is a logical variable that describes whether an advertiser is allowed the impression (ie, the advertiser has won with its bid). It is clear that click is directly dependent upon imp (ie, click cannot be true if imp is false). Therefore, we use only bids where imp == True as training and test data.

In our dataset, Browser indicates the user agent of the user targeted by the bid. We implement a simple parser to map Browser to two variables: OS indicates the operating system (Windows, Mac OS X, Linux and other) and Browser indicates the browser family of the user (Firefox, Chrome, Safari, Opera, Internet Explorer and other). AdvertiserID contains the unique identification number for each advertiser. To reduce our memory overhead, we will use only bids where AdvertiserID == 2821 (footware seller).

The actual user-identifiable variables are the interest_*-variables. They indicate whether the user has shown a specific interest into a tag used by iPinYou. For example, if the user is known to be interested in education-related material, interest_education == True. The same holds true for the Inmarket_* and Demographic_gender_female variables. We remove the redundant Demographic_gender_male variable.

Notice that there are additional variables in the dataset, which we chose to ignore (at least for the initial analysis). Some of them are completely irrelevant to the click prediction, such as the domain name of the ad. Other variables are discouraged by the distributor of the dataset, because they are not consistent. This holds true for the conv variable, which indicates whether a user actually converted to a customer after clicking on an ad.

The following table lists the discarded variables:

Name	Type	Reason
Index	int	Obvious
BidID	int	Identifier useless for classification (different for every observation)
Time_Bid	int	Bids were conducted only on one afternoon - results wouldn't generalize
UserID	int	See BidID
IP	string	Useless on its own, ISP data has no connection to ads, location data can be extracted, but is already present in Region
Domain	string	Variable is hashed, referral page of ad probably doesn't exist anymore and has only post-click effects
URL	string	See Domain
AdslotID	int	See BidID
Adslotfloorprice	int	Information is already contained in Payingprice
CreativeID	int	Variable is hashed and very specific to the ad
Biddingprice	int	Merged into Payingprice
conv	int	Variable affects only post-click period and is discouraged by provider due to the non-standard conversion definitions of advertisers
Region	int	Doesn't generalize beyond Chinese market
City	int	See Region
Demographic_gender_male	bool	Negative Demographic_gender_female

Classification

To predict whether a user will click or not is a classical binary classification task. Data details that we have a binary label and binary/categorical features. We figure that to regress click ~ ., random forests are a perfect fit. They are specifically designed for categorical features. Random forests are ensembles of decision trees and therefore reduce the variance of single decision tree algorithms.

We also analyze the dataset using a logistic regression. This serves as a robust baseline.

We will use scikit-learn and statsmodels to implement the algorithms.

Usage

Notebook

Place the ads.csv in the data directory. Start jupyter notebook and open click.ipynb. Follow all steps.

Scripts

Install dependencies with:

pip install -r requirements.txt

Place the ads.csv in the data directory. Then run ./prepare.py to clean the dataset and prepare it for training. It will output data/ads_clean.csv. Use ./plot.py to plot a correlation matrix. Use ./train.py to train a random forest or ./train_lr.py to fit a logistic regression. The help pages display additional information.

Further considerations

As seen in Data, there is a variable imp that describes whether a bid from an advertiser has been successful. It would be interesting to regress imp using the other variables. However, we figure that the provided variables are not conclusive enough to classify imp correctly.