How AI decides the price of 80 billion online auctions - a day
The convergence of marketing, artificial intelligence and machine learning is creating a whole new breed of buyers holding up virtual signs at up to 80 billion daily auctions.
At stake is your attention – and potentially your wallet, as the auctions are for potential customers.
The auctions start every time you visit a homepage, and are about the ads you see – with the buyers on the other end probably knowing a lot more about you than you think.
This is in large part due to tools like artificial intelligence, and to leverage these new advantages companies need to learn how to use them or turn to experts that specialise in this area.
One such expert company is Rocket Fuel. Since its beginning in 2008, Rocket Fuel has offered clients the possibility of targeting online advertisements to customers in a way that relies on smart systems making informed decisions about buying add space in less than a tenth of a second – and do so up to 80 billion times a day. The company has been so successful at this that it currently sits comfortably in the top 20 of Deloitte’s list of the 500 fastest growing technology companies in America.
At the heart of the company’s engine room, you find Mark Torrance, the company’s Chief Technical Officer. From a background of studying AI technologies at Stanford University, and applying that knowledge to various practical solutions in fields including finance, email assistants, and intelligent data visualization, he has a strong understanding of what AI and machine learning has to offer. I spoke to him about his take on AI and machine learning techniques for ad buying and what possibilities, trends and challenges he sees on the horizon.
Jakob Sand (JS): To start with a bird’s eye view, could you describe the influence of AI and machine learning on marketing, and how is it applied?
Mark Torrance (MT): I would say that AI is about having computers systems or software do things that require reasoning on their own. If we think back 50 years, AI was about making computers better than humans at playing chess or checkers. We have long surpassed that and recently added Jeopardy to the list of things computers can do better than people.
Today, we find that AI and machine learning is being applied to many areas and industries, especially those that have abundant amounts of data available.
One of the reasons for this is that we can apply machine learning, which involves software systems that become better at performing their tasks over time through a trial and error approach.
An example is the Roomba vacuum cleaning robot that gradually builds up a map of your rooms and identifies the best way of cleaning them. Machine learning was also the basis of the Google self-driving car, where it in part taught itself how to drive safely by logging thousands of hours on the road and slowly improving.
To programme that sort of capability into a machine would be a nearly impossible task, and it is only thanks to machine learning that it was possible to develop the now locked software that runs a Google Car.
Marketing is also amenable to machine learning and AI, because large parts of the field is data driven when you think about it. This includes patterns of behaviour that could suggest that a person is interested in buying a product. However, identifying such fine-grained patterns has been impossible until recently.
One reason we can now do it is that computers have gotten much faster, and able to compute the data we have available at greater speeds. A second reason is the fact that we have much more data now.
At the same time, the places where people are spending their time and attention are also becoming digital in the form of online webpages and apps. These platforms give us information about individual user activities that we did not previously have, and this has led to the current convergence of AI, machine learning and marketing.
JS: What are some of the things that you can do with AI and machine learning that you cannot do with traditional marketing tools?
MT: At the end of the day, machine learning is about improving, and one thing that I think is important to stress is that these tools offer you just that – marketing performance that improves over time.
A traditional marketing campaign starts with identifying what sub groups of customers a campaign will be targeted towards. Once these groups are identified, the ads are then aimed at the whole group through various channels, like TV adds, radio spots, or wholesale buying of advertisement space on websites.
AI and machine learning lets you work with a much finer granularity when it comes to sub-grouping consumers, sometimes right down to the individual level.
We do this by using data such as what the customer is looking at right now, geographic location, language settings, whether they are using a PC; laptop, MAC or mobile device and what sites they visited before the one they are currently on. It might potentially include what they have previously shown an interest in, and how that might relate to the current site they are visiting.
JS: This better granularity of customers and precise targeting is only half the story here, is it not? Staying with time, they also let you make decisions at a speed that no human can compete with?
MT: Yes, that is true. Every time a customer visits a news website, checks in on Facebook or YouTube or any other kind of website that has advertising space, it automatically starts a bidding auction where companies compete to show their ads in that space.
The auction is concluded in 100 milliseconds, or one tenth of a second. In that time our AI looks at the data for this specific website, the person visiting it, what data we have available about that person, and what campaigns we are running that might fit his or her profile, as well as deciding what the maximum price is that we are willing to pay for that specific advertisement space. This all happens in a tenth of a second, and we take part in 80 billion such auctions every day.
One of the ways we do this at the backend is to correlate data over time. We have built proprietary software on top of the open source platform Hadoop; our software looks for specific patterns in consumer behaviour and uses them to create predictive models of the customers.
These can include adjusting campaigns to user groups who seem to be going through major changes in their life, based on the data we have on them, or it might be basic information that suggest that they have certain behaviours at certain times of day that might make them more likely to respond to particular advertising campaigns.
The adaptive nature of algorithms built on machine learning means that these models can adapt campaigns to changes in consumer behaviour and that we can wave goodbye to static, simple consumer models and say hello to models that adjust themselves automatically on the short, middle and long term.
JS: While the software and adaptive nature of machine learning gives you new possibilities, there must also be issues or challenges? What are some of the major challenges you face in coming years? For example, I could imagine that gathering data is a double-edged sword. On one hand, it gives you a better foundation for making decisions, but at the same time, it also pushes both hardware and software.
MT: I would agree that data is an interesting dilemma. In the past Rocket Fuel has perhaps erred on the side of caution and saved as much data as we possibly could. It does generate a situation where data scientists keep asking for more and more data whereas the guys in charge of infrastructure are asking ‘do you really need to save all of it?’ The particular selection of data and what might be considered valuable is definitely something that presents a challenge. Due to the size of investments in hardware a company needs to make to save data, we are forced to become selective – especially as we see more and more new data sources adding information to the mix.
Another major challenge is attribution. When a person buys something after seeing ads, was it because of the ads they saw, or would that person have bought that product anyway? The people might also have been exposed to the same campaign in different contexts – for example online and on TV – which raises the question of which part of the campaign worked best. There are several companies trying to solve this “attribution” problem and provide tools to marketers, but I think it is an area that needs to mature.
JS: Looking under the hood of AI and machine learning brings us face to face with algorithms. Will the future of marketing be based around competing algorithms that are locked in a battle to be the best at predicting what will make customers buy something or react positively to a specific campaign? Or will there still be space for the human element in online advertising, when everything is a battle of making decisions in milliseconds?
MT: Well, for all the encroaching that you could say that systems based on AI and machine learning are having on marketing, we are still some ways off in supplanting the human element.
In pop music, machine learning and AI have been used to predict what songs will become hits if they receive enough airplay, but even in that domain most decisions continue to be made by people.
So far, we do not know how to design an algorithm that can create the content of an ad campaign that will help ensure it “goes viral.” You can say that we are in a situation where AI and machine learning can tell you who to show ads to, but they are not yet as helpful at determining what the ads and messages should depict. The systems do not have the creative knack for knowing what will make a person work as a call to action.
One thing you could say is that software like we have at Rocket Fuel frees up a lot of the creative potential in marketing and lets people go and do what many of them got into marketing for: to use their creativity to find ways to engage with and influence consumers.