machine learning

Digital Advertising: A Hotbed for Machine Learning

Professional headshot of Anand Das
By Anand Das, Cofounder & CTO
November 21, 2017

The advertising world is getting more complex each day. Access to the Internet is now pervasive across a wide variety of devices in most places around the globe.  So advertisers must now consider user preferences more carefully than ever to reach their target audiences. All of this leads to further complexities since it requires analyzing large data sets with sparse signals about their preferences and needs.

With ever-changing algorithms to drive revenue for publishers and ROI for advertisers, both sides of the ecosystem need to continuously evolve. That’s why I think that advertising is currently a hotbed for the application of Machine Learning (ML).

What machine learning does

ML analyzes large data sets: tens or hundreds of billions of ad impressions, hundreds of millions of users and thousands of buyers and advertisers. It works to identify trends online in real-time. On the buy side, the insights garnered allow advertisers to act on issues and opportunities, such as ad blindness or relevance, by optimizing cost-per-action (CPA) and click-through-rates (CTR).

Likewise, ML can also be used to solve complex problems on the publisher side. Let’s go through some of them:

Publisher revenue optimization:

In the past, the old waterfall system was determined by manually set prioritization rules, based on aggregate historic pricing data. Auction dynamics changed with the advent of header bidding and the swift adoption of wrapper solutions, like PubMatic’s OpenWrap. Platform data and ML can be used to predict how advertisers will value an ad impression. Then a price floor is set in the auction using ML based on the interest level from buyers and then the floor price is submitted to the bidders.

From here, you can compare trends to real-time data and analytics to learn additional insights that help to determine how to set future floor prices. The publisher is then able to reclaim more control over the ultimate price, effectively increasing yield.

Leveraging ML for pricing helps protect the value of a publisher’s inventory and alleviates concerns regarding appropriate value capture from buyers.

Managing ad quality:

Ad quality is critical for publishers and is only growing in importance. The rise of programmatic has made it more difficult for publishers to maintain quality control over advertisers, ad units and creatives served on their sites. Publishers need to balance monetization with site performance to offer a high-quality user experience.

PubMatic has developed ad quality tools leveraging ML which can help predict if an ad is likely to be “bad” before it shows up on publishers’ sites. This allows us to take action pre-emptively.

For example, we scan ads as they come through our platform and use ML to classify an ad as harmful (malware, objectionable content, high bandwidth, etc.). Should we find one, we block it before it’s shown on the publisher’s site.

To keep improving, we look at hundreds of ad variables and overlay known “bad” creatives to feed our algorithms. ML, over time, learns which partners and ads are the least trustworthy and blocks accordingly.

Using ML helps identify and block ads with the varying partner definitions of a “bad” ad. PubMatic also looks at which DSPs or buyers have this type of creative, allowing us to block based on historical trends. In this way, ML also helps with brand safety.

Managing infrastructure costs by implementing bid throttling:

Volume and pixel management have become key issues with the advent of header bidding. The average sell-side platform (SSP) and demand-side platform (DSP) are receiving more traffic while filling smaller percentages of inventory.

DSPs are assessing whether an opportunity exists to buy through the most optimal set of pipes to help minimize queries per second (QPS) and infrastructure costs. So how do we get more out of the given infrastructure to deal with the QPS challenge?

SSPs are playing their part to minimize QPS with activities like bid throttling. Throttling leverages ML to allow the SSP to only send traffic to DSPs that are likely to be interested in that traffic. Consequently, this increases the efficiency of integrations between platforms.

In order to mitigate the spiraling infrastructure costs, we use ML to predict which DSP is likely to bid and/or win. We identify important attributes that affect a DSP’s bid and win rate, create weightings for the attributes identified, and learn from each buyer’s temporal behavior to improve our predictions over time.

If you can predict who is likely to bid for and/or win an impression, and then send that impression to the DSPs most likely to win, you can generate significantly fewer bid requests.

By leveraging ML for bid throttling, we are able to make DSPs’ businesses more efficient. Additionally, by being a more efficient technology partner, we are able to pass that transaction processing savings on to our clients.

What’s next?

The examples above are just a few of the areas where ML is being applied within PubMatic and across the advertising industry. Contact us for additional information and discover how we can partner with you.