Display online advertising is a multibillion-dollar industry. It comes in different formats, including banner ads composed of text, images, flash, and rich media such as audio and video. Advertisers or their agencies place advertisements on a variety of websites, even mobile apps, across the internet to reach potential customers and deliver an advertising message.

Display online advertising has served as one of the greatest examples of machine learning utilization. Obviously, advertisers and consumers are keenly interested in well-targeted ads. The industry has relied heavily on the ability of machine learning models to predict the effectiveness of ad targeting: how likely it is that an audience in a certain age group will be interested in this product, customers with a certain household income will purchase this product after seeing the ad, frequent sports site visitors will spend more time reading this ad, and so on. The most common measurement of effectiveness is the click-through rate (CTR), which is the ratio of clicks on a specific ad to its total number of views. The higher the CTR in general, the better targeted an ad is, and the more successful an online advertising campaign is.  

Click-through prediction entails both promise and challenges for machine learning. It mainly involves binary classification of whether a given ad on a given page (or app) will be clicked by a given user, with predictive features from the following three aspects:

• Ad content and information (category, position, text, format, and so on)
• Page content and publisher information (category, context, domain, and so on)
• User information (age, gender, location, income, interests, search history, browsing history, device, and so on)

Suppose we as an agency are operating ads on behalf of several advertisers, and our job is to place the right ads for the right audience. With an existing dataset in hand (the following small chunk is an example; the number of predictive features can easily go into the thousands in reality) taken from millions records of campaigns running a month ago (let's say), we need to develop a classification model to learn and predict future ad placement outcomes: