Cognitiv: How ‘Neural Networks’ Could Be The Key To Improving Programmatic Placements
The company is helping marketers make sense of big data through advanced machine learning.
The holy grail for marketers is to be able to reach the right consumer, with the right ad, at the right time — but, of course, not all consumers are created equal. And neither, says Cognitiv CEO & co-founder Jeremy Fain, are programmatic placements.
“The first solution we built was programmatic,” Fain explained. “We can take all the historical results of [a marketer’s] campaigns, and then we optimize a neural network to predict an outcome. Then we input that probability into a bid creator, and it creates [a price of] a dollar, or 10 cents, or 50 cents. Then, we can put that into a DSP and buy media. Basically, every single impression gets a different bid based on the inherent value our neural network predicts,” ideally making a marketer’s spend more effective and driving better results.
Neural networks? Yes. By using this brain-like, sophisticated form of machine learning, Fain posited that Cognitiv is able to help companies place bids more effectively and drive ROI in the long run. Here’s how.
GeoMarketing: So, what are “neural networks,” and how does Cognitiv harness their power to improve programmatic ad placements for brands?
Jeremy Fain: Well, with programmatic, you have all the bid stream stuff coming in: all the content, or location, or apps. Then you have all of the additional third party data, all the first party data — you put all of that together, and you have this enormous, complicated data set, and really nobody had come up with a great way to make sense of it.
So, you have all of this data, and everybody has it. But it is not all compiled together, and it is not being used together. That’s where [we] come in: The beauty of neural networks is that they work like the human brain. It’s basically the most advanced pattern recognition machine in the known universe.
Computers have historically have been able to do things that you tell them to do: You program them to know that two plus two equals four, and they can do it a trillion times. But if you ever ask them to recognize your face and then you move your face slightly, [the computer] couldn’t do it. It could only recognize data in exactly in the same way.
So what computer science has been trying to do for forty years is program computers to think the way the human brain does it: through pattern recognition. With neural networks, it’s about training computers to learn and recognize patterns the same way that we do. Then, the added advantage of our official neural networks is that they can take much, much more complicated data that we as humans can understand consciously and find patterns in all of it.
How does that help marketers improve their ad placements and targeting — versus what they could do organically based on classic demographic data, et cetera?
Here’s an example: Take, say, Trident gum. Basically everybody in the whole world might theoretically want to buy Trident gum, but as a marketer I have to make a decision right now to say “I am only going to market to males,” or to this age group, or to this income level in this area. There are over 300 million people in the United States, and maybe a full 50 million that buy gum — but which 50 million? Who knows why they do it, because the reasons vary more than you might think.
Those 50 million people that buy gum are not all within a certain age range, or have visited a certain place, et cetera. There are infinite reasons why people buy gum, is what I’m getting at.
A neural network is able to essentially create an infinite number of clusters and fill in all of those gaps that one or two big overlapping clusters can’t. What [Cognitiv] does is take that cluster and break it apart so that marketers aren’t actually wasting 25-30 percent of people in there that really don’t fit.
That is why neural networks are so much more advanced than other machine learning capabilities out there: They are what we call universal function approximation. If the function is people that buy gum, and if you give it enough information, it will find the exact 20,000 people who do.
To sum up: We know that in digital marketing these days, there is simply a ton of data. We take that big data, we latch on our platform which is called Neuralmind. Then, what we like to say these days is, we “weaponize” that data.
How? The first solution we built was programmatic. It can take all the historical results of [a marketer’s] campaigns, and then we optimize a neural network to predict an outcome — conversion, click through, video completion, et cetera. Then we input that probability into a bid creator, and it creates [a price of] a dollar, or 10 cents, or 50 cents. We can put that into a DSP and buy media.
Basically, every single impression gets a different bid based on what our neural network predicts. There is [a different value] for that impression, this user, that content, that instant.
So the goal is to to place bids more efficiently to drive ROI.
Right. Currently, one DSP will use the same algorithm for, say, GM trucks and Kellogg cereal. They have different audiences that they choose from in how they decide to bid, but all the information that’s coming in is basically used the same way.
What we can do is to build out, out of all the people that came to a dealership, we can say who came to sign up for a test drive. We start there, and then we move out and figure out how to value those people and those impressions [differently]. And then everybody who clipped a coupon for Kellogg, same thing.
The key is that we don’t have to go in with a hypothesis about who those “coupon cutters” are going to be; it’s important not to assume. We aren’t only going to target moms — we are going to target who will actually be most effective, and we are going to dynamically value every single impression.
Both historical and real-time location data can be great predictors of consumer mindset. How does location play into what you do?
It is just one input, but is is a very important input because it provides a very rich view of humanity. It is a very rich view of your behavior, of where you go, of what you do; it [tells us a lot] if we know that someone searched for something on their phone and then that phone was carried into Best Buy, or T-Mobile, or somewhere similar.
Anyway, if we can follow this one user’s pathway and then compare those pathways to all the users, it is only going to give us better outcomes. Understanding what people are doing offline, to a certain level, and then being able add it to all the online data is a huge step forward.