A Machine Learning approach to real-time display advertising

Click here – it’s not an ad, we promise ;-)

Until recently we relied on a batch-based algorithm named Compass that assigned exactly one campaign to a user or a website every two hours. This assignment was then used to reply to real-time bid requests. You can read about how Compass did this in detail in a previous blogpost.

However, we encountered limitations to Compass's capabilities. While display advertising is most efficient with fast decisions in real time, Compass wasn’t flexible enough to deal with market changes. For example, if we assigned one campaign to a certain user pool (or website pool) as a batch, those users would have not been able to see any other ads within the next two hours. Or, if the assigned campaign didn't have an ad for the currently requested format, the user would simply not receive an ad. On top of this, specific campaigns could struggle with spending their budgets efficiently if less traffic than expected was generated in assigned auctions.

So we needed a system that provides rapid decisions for bid requests.

Figure 1: RaBid – bidding process overview

Looks super easy, doesn’t it? To explain it in a little more detail, let’s take a closer look at a bid request containing user and website information. As previously mentioned, the goal is now to decide which campaign is the best choice (e.g., in terms of click probability) and how much we are willing to pay for that auction.

We start by scoring all suitable campaigns for this incoming bid request. We then assign advertising budget to the best campaign based on the remaining budget for each campaign and the calculated campaign score. After that, we can calculate the bid depending on campaign score and market context (e.g., expected competition). With this bid we can now participate in an auction. Our ad is displayed if we have placed the highest bid; otherwise our bid is discarded – but that still gives us a valuable market insight.

 
Let us offer an example and go even further into detail. OTTO is a leading e-commerce platform, where you can basically buy anything you want. As an online shop, different marketing methods are used to attract customers – for instance, display advertising. With this method we have different campaigns with different contents. Let’s say we have three campaigns: one carrot campaign, one apple campaign, and one t-shirt campaign. One of our customers is Berta Bunny, a frequent visitor to otto.de. She is highly interested in carrots, sometimes also in apples. Berta is a proud user of her recently purchased smartphone and spends her evenings reading about the newest vegan cooking trends.

Figure 2: RaBid – Berta Bunny’s journey

As mentioned, the bid request we receive contains user information (if available), and context information such as page URL and ad size. Furthermore, Orbidder(our SSP/DSP solution) enriches the bid request with information on all currently relevant campaigns before forwarding it to RaBid.
Remember Berta Bunny on the couch browsing for new recipes? As she opens her favourite food blog, a bid request is triggered. It contains her cookie ID and the page URL. The request is enriched with the information about our three currently active campaigns and their expenditure so far.

But how do we now decide which campaign should be displayed?

We start by scoring all active campaigns suitable for that bid request. To do this, we will look up available user features, for instance affinity and activity profiles for Berta Bunny. Additionally, we will consider context information such as website click-rate and category (maybe it's a fruit shop... or a fashion blog?). We also want to include our knowledge about the products advertised by the active campaigns and so we include campaign-specific features.

All these features are used in a neutral network to predict different KPIs such as click probability. In each bid request a score is calculated for every valid campaign. To be able to compare the scores across these different campaigns we normalize them. Above-average scores are rewarded and more likely to be chosen.

In our example, this means that Berta Bunny’s affinity and activity profile indicate a high interest in carrots and apples. She is currently browsing a food blog on her smartphone – and we know this food blog is highly frequented by other OTTO users who sometimes click on our ads, preferably on food content. This means that the food blog is an attractive website to place OTTO advertisements. Based on all this information we score each of our three suitable campaigns. Because we have two campaigns in the food segment, the carrot and the apple campaigns will most likely receive a higher normalised score than the t-shirt campaign.

The normalized score is then used for the campaign assignment.

Besides the normalised scores, the remaining budget is the main factor for assigning campaigns and is updated in real time. All suitable campaigns are compared based on their relative expenditure. Campaigns with a low percentage of budget expenditure are pushed in the assignment process to achieve similar budget ratios.

 
Let’s get back to Berta Bunny. Although Berta has a strong affinity for carrots, she also likes apples a lot! Now, the apple campaign has had trouble spending its budget during the day, whereas the carrot campaign budget is well on track. The apple campaign is therefore pushed in the assignment and selected for the bid calculation.

To calculate the final bid, we need the normalised score of the assigned campaign, information on the bid function, and a controller factor. The bid function is a function that assigns a bid value to a given KPI score, depending on the website context. It reflects the competition in the auction process; the stronger the competition, the more we need to bid to have a higher probability of winning.

 
Now, we could insert the normalised KPI into the bid function and get our final bid. However, we need to make sure that all campaigns spend their whole budget during the day. Since all our calculations are based on predictions, we are constantly at risk of over or underspending. Therefore, we have a controlling component to scale the score and so adjust the bid value.

In our example, we consider a popular vegan food blog with strong competition among advertisers. Additionally, Berta’s score for our apple campaign is high and the controller needs to compensate for the underspend. In conclusion, all three factors lead to a high bid and hence to a high winning probability.

The process for all components – scoring, assignment and bid calculation – happens within approximately 20 milliseconds. With the resulting campaign and bid, we participate in an auction with other competitors. The highest bid submitted wins and the winning advertiser can display the campaign.

Given our high bid for Berta, we outbid our competitors and won the auction to display our apple campaign on the food blog.