You’ve probably heard a lot about machine learning in recent years… but what does this term really mean? Well, if you’re buying media programmatically, it’s the basis of everything you do – and it could just be your secret weapon.

ML

Did you know that, according to Google Trends, interest in machine learning has rocketed by over 600% since 2014? With innovations like smart assistants and self-driving cars, that may come as no surprise.

But, outside of consumer products, how does this technology help media buyers get more out of their programmatic campaigns in business? And could data science really help computers make better decisions than a living, breathing human being?

At IPONWEB, we’re proud to have built some of the world’s first programmatic systems and the algorithms that power them, so we’re well-placed to answer these questions.

With that in mind, let’s put machine learning under the microscope, discover how it compares to human-based media buying – and why it’s much more than simply a buzzword.

Table of contents

  1. What machine learning really means in programmatic
  2. Human-powered programmatic optimisation
  3. How machine learning can help media buyers
  4. Machine learning and programmatic media buying in the real world
  5. The hybrid future of programmatic advertising

What machine learning really means in programmatic

 

Long before inventory could be bought and sold in real-time – 200ms or less – it was all about direct buys. Phones were being dialled, hands were being shaken, and deals were being done the old-fashioned way.

With the advent of Real-Time Bidding (RTB), brands and agencies were able to go far beyond what was possible with traditional media buying.

No matter how much a fast-talker a buyer might be, there’s a hard limit to how many deals can be done over the phone. Not only that, but they were also limited by how many different inventory suppliers (and, by extension, audiences) they had access to.

With RTB, all of this changed.

Why? Because of the sheer volume of inventory this technology made accessible to media buyers. All of a sudden, there was so much inventory available across so many different publishers that no human on earth would be able to parse it all to make truly informed buying decisions. Remember when TV channels grew from a small number of national broadcasters to many different cable operators? The move to RTB is similar, but on a much larger order of magnitude. This explosive expansion in programmatic meant that the floodgates were open and digital media buyers were suddenly challenged to find their target audience from a potential pool of millions.

Faced with this never-ending firehouse of opportunity, buyers needed a way to make split-second, data-driven decisions reliably without simply discarding potential opportunities at random.

It’s here that machine learning based algorithms started to evolve – and fast.

The algorithmic approach to media buying meant that users could near-instantly decide on the supply that best suited their goals, and what to bid for those opportunities, based on hugely expanded criteria, including:

  • First-party data. This is the data you’ve collected about your audience from website actions, CRM systems, social media, and more.
  • Campaign data. This means using the performance of past campaigns to inform new ones. This concept is central to the way machine learning works: taking what’s been successful in the past and using it to secure more success in the future.
  • Bidstream signals. These signals, or metadata, found within the bidstream might be contextual, demographic, geographical, or even relate to the universal identifiers which are paving the way for the cookieless future.

The use of data in this way also broke a long-established link between the buyer and the media partner. Rather than relying on the publisher to advise buyers of how best to reach an audience, RTB and machine learning meant buyers could leverage their own data to take control of their outcomes more directly than ever before.

Human-powered programmatic optimisation

 

While machines might have the edge in terms of speed (and a few other things, which we’ll cover shortly), we humans still bring a lot of value to media buying:

  • We are able to process and take action on client briefs, meaning the actual planning of a campaign can only be done by a human. In addition, the ability to plan media buys based on intuition and experience will always be critical to a well-executed campaign.
  • We can leverage more fine-grain tools based on external factors, such as inclusion / exclusion lists and manual bid multipliers to enhance the strategies defined in the media plan. It’s here that factors such as emotion and human sentiment come into play. For example, there may be times – during natural disasters, for example – when running certain ads wouldn’t be appropriate. This is something machines will likely never understand.
  • Humans are (for now) far better equipped to make decisions about brand perception and consumer impact. For example, if there are certain real-world factors you don’t want your brand to be associated with, making those calls still very much requires human input.

Beyond the nuts and bolts, we can also bring our creativity into the mix with creative design, copy, and the big ideas that power large-scale campaigns – something machines are quite some way from achieving.

So, what about the drawbacks of being human when you’re buying media?

To better illustrate them, let’s use a practical example based on a media buyer’s everyday life: managing line items.

Without the assistance of a computer or machine learning algorithm, you’d need to manually assess each of the line items you create. You might have a database of historical campaign performance, from which you can pull key data points to use for optimisation: audiences, time of day, creative sizes, supply source, and so on. For each line item you create, you’d need to cross-reference historical data and leverage those insights to make your planning decisions.

All of this takes time. A lot of time.

In fact, even the best data scientist in the world with the most comprehensive dataset on the planet wouldn’t be able to compete with a machine, because humans have a hard limit of how much we can process – and how fast we can act on that data.

The scope of our decision-making is also limited by the range of the dataset we can manually parse. For example, you might see a strong correlation between two line items and the audiences they were assigned: audience A delivered better results than audience B. Seems straightforward enough.

But what if there are other correlations you’re just not looking at, or don’t have the bandwidth to analyse? For example, a human may not notice that a particular campaign is performing well with under-30’s in the New York area who are using iOS devices, but a machine – with its huge data processing capabilities – almost certainly would.

How machine learning can help media buyers

Now that we’re familiar with what people can bring to the table, what about machines?

Machine Learning – computer systems that are able to learn and adapt without following explicit instructions, by using algorithms and statistical models to analyse and draw inferences from data sets  – opens the door to a host of benefits which, in many ways, exceed the capabilities of even the smartest of us.

Here’s how machine learning could benefit your campaigns:

  • Algorithms can make thousands of decisions per second with – crucially – no second-guessing. Subjectivity is left at the door, because all media buying decisions are made by the data – for good or for bad.
  • Machine learning can instantly call on huge volumes of historical data from previous campaigns. What worked? What didn’t? This data can all be factored in before a media buy is made.
  • Because algorithms get better the more data they have to pull from, machine-based decisioning offers continuous improvement and optimization.
  • The ability to launch largely automated campaigns with minimal human intervention means machine learning helps advertisers achieve speed and scale. Buyers can simply replicate previously successful campaigns, tweak a few parameters, and go live in moments.

Machine learning and programmatic media buying in the real world

So far, we’ve looked at the theory of machine learning in quite some detail… but what about the more practical side? What does machine learning look like in the day-to-day world of programmatic?

To answer that question, let’s look at a couple of examples of how algorithmic techniques are used to buy – and optimise for – programmatic media.

  • Traffic filtering. Through sheer force of numbers, it would be practically impossible to listen to every request in the bidstream. Our programmatic infrastructure layer, BidSwitch, already sees over a trillion requests per day – and that number is growing all the time. To manage this volume, machine learning based algorithms can carry out traffic filtering (or shaping) based on specific buying patterns to ensure you only listen to the supply you’re most likely to buy.
  • Testing, testing, testing. As the name suggests: machine learning is about trying new things to see what works. Unlike a single human, an algorithm is able to try new things over time (such as new geographies, audiences, times of day), then learn from these tests to further optimise performance based on what works – all in real time.
  • Fraud detection. As a means to protect buyers from unwanted media waste on fraudulent opportunities, a machine learning algorithm can study and detect ad fraud as it happens, cutting off this activity at source before it becomes a big problem. Even better, it can use what it learns to prevent further interactions with fraudulent supply.
  • Streamlined campaign optimization. Machine learning enables out-of-the-box buying strategies which media buyers can apply in a few clicks. These strategies can be used to optimise buy decisioning, including both specific supply and bid amounts, based on specific criteria. Preset buying strategies can optimise to specific goals – clicks, conversions, win rates, etc. – within wider constraints, like maximum CPMs.

The bottom line here is that, for advanced media buyers equipped with the right tools and customisations, machine learning is a powerful tool that can be used to drive outcomes specifically tailored to their performance goals.

These customisations make it possible to create entirely bespoke traffic shaping, campaign targeting, and buying optimization strategies which leverage far more advanced criteria, like viewability, 1st-party identifiers, ad unit interactions, video view time, and much more. The key here is that these strategies can be designed entirely around you, making them a much more powerful tool to have in your arsenal.

Want to know more about how to craft bespoke buying strategies to unlock your programmatic potential? Check out our free downloadable guide for the complete picture.

The hybrid future of programmatic advertising

Now that we have a clearer vision of the part machine learning plays in this ecosystem, where does the future of programmatic media buying really lie?

For us, it can be summed up in just one word: hybridity.

It makes little sense to cast aside the benefits of human input – creativity, intuition, logic, emotion, goal-setting – for a pure-play focus on an algorithmic approach. Machine learning is a powerful tool, but it should be used in combination with everything people can bring to the table.

So, in a perfectly balanced programmatic process, it’s the humans that will do the planning, and the machines that will execute on it with lightning-fast real-time decisioning to supercharge campaign results.

After all, teamwork makes the dream work.

Want to learn more about how you can supercharge your media buying by pairing machine learning with the human touch? Download our free handout and discover how easy it is to build custom buying strategies to unleash your programmatic potential.