Data Collaboration Platform Playbook On-Demand Get Yours Now

Modeling is Harder Than it Looks: Q&A with Chief Data Scientist Omar Abdala

Modeling is Harder than it looks: A look at Lotame's Audience Optimizer Tool

The rising cost of customer acquisition is a marketer’s migraine. Brands are spending 60 percent more to fill their funnel than they did just five years ago. And the buck doesn’t stop there. Retention costs are also continuing to climb, with $148 billion spent on global loyalty programs — all for a hard-won, yet very easily lost asset. 

Recognizing the high stakes facing brand marketers, Lotame’s data science team set out to make our Audience Optimizer tool smarter, faster, and easier to use. With the evolution of Audience Optimizer, comes a more efficient tool for brands to supercharge their prospecting strategies, and win those next best customers.  

So what exactly is Audience Optimizer and how can it help you fill your funnel? Our tool extends your valuable first-party data using powerful AI/machine learning and the largest dataset available outside the walled gardens. The result is Data Empowered and actionable modeled audiences with the right balance of precision and scale.

Yes, it’s a mouthful. That’s why we sat down with the “math man” behind the tool himself, Chief Data Scientist, Omar Abdala, to learn in his own words why Lotame’s Audience Optimizer is the smarter, faster, and easier way to crush customer acquisition. 

Tell us about Lotame’s audience modeling capabilities and the Audience Optimizer tool. What’s first, best, and most about it?

When we talk about machine learning models, there’s a fair amount of misconception and mythmaking. Our audience model is really nothing like what marketers are imagining.

Basically, we take a list of 10s of billions of profiles and try to figure out, based on the characteristics of a marketer’s first-party data (seed set), where do the rest of those say 20 billion profiles rank. That’s a very substantial piece of big data engineering that’s done, and that’s really the unique thing about our tool. The magnitude of the scoring, ranking, and modeling performed over one of the world’s largest sets of third-party data — it’s an outstanding achievement, of which we’re immensely proud. Humans are incapable of this kind of activity with the amount of precision and scale we’ve achieved.

Can you tell us about some of your AHA moments during this work?  

There was a lot of work done on the algorithm optimization, but there are two things that really stand out. The first was being able to efficiently score billions and billions of profiles. Typically people think of modeling in terms of training off the seed set but there’s a much more significant contribution from efficiently scoring those billions of profiles. Data science, engineering and the technology teams make this type of magic happen. That’s the major innovation, and what would be extremely difficult for any other group to replicate without investing some herculean effort and resources. 

The second critical discovery we made which really makes a huge impact on the quality of our models came from our work with Data Stream. For example, on any one profile you may have behaviors 1, 2, 3, 4. On a surface level, it’s easy to say “Oh, I see behavior 1 so it’s checked or unchecked.” That’s not really the case. Because we sync with data providers, there’s a chance we get to see the cadre of behaviors or we don’t. Why is this important? It’s not an easy split: “Have this behavior / don’t have this behavior.” There’s a third option that exists in which “Don’t have this behavior because data provider doesn’t provide behavior. It’s a distinction that gives us an additional layer to compare behaviors to, which provides a more accurate profile in return. 

What are the biggest misconceptions around audience modeling in the industry today?

When it comes to audience modeling, and AI in general, there’s this sense that something magical is happening. For those of us who are practitioners in this space, we get it. There are cost functions and optimization routines. It’s all math and linear algebra for us. But for others, there’s this thought that you put in a question, and the machine learning model just mysteriously produces the right answer, always. 

The big misconception is that once you’re using machine learning, everything has to make sense. The reality is that it’s very dependent upon what you put in and the algorithms are very effective at picking up what you tell them. If you don’t articulate exactly what you mean in that seed set, it will pick up on something you didn’t mean. 

That can be intimidating for marketers not trained in this space. In your opinion, do you need a data science degree to use Audience Optimizer effectively? 

No, you definitely don’t. Our objective was to create a tool that could be used by marketers, analysts, data scientists, etc. The most important criteria is having a strong sense of what your objectives are and what you’re looking for and looking to avoid, in terms of metrics or profiles. 

Can a marketer, who may not have a ton of first-party data, benefit from audience modeling in general and Lotame’s tools in particular? How? 

Sure. Let’s take a look at a CPG example. Say a cereal brand wants to figure out, out of those who purchase cereal, who is most interested in sugary, chocolatey cereals. 

Audience Optimizer can analyze survey responses to pick out respondents in a large set of third-party data, and determine the difference between someone who buys whole grain cereal, and someone who buys the sugary stuff. Using our tool, you’re able to set “cereal buyers” in the background, and “sugary cereal buyers” in the foreground. The model then picks up the difference between a general cereal buyer and a chocolatey, sugary cereal buyer — all using third-party data.  

What about brands who collect a ton of first-party data? How can they utilize our tool? 

Let’s take a look at a retail media network. They know how to target cereal buyers, unlike CPG, as they have a wealth of purchase data so their model can really push the envelope. The actions they collect from their own footprint are closer to a purchase than we typically see. This allows for more advanced conversion. By constructing a background of profiles from a brand’s campaign in-channel, the retail media network can track all the way from purchase back to the marketing activity that led them there. This allows for more definitive optimization around the sale. For example, they could set “dollars spent on product” as their objective versus “dollars spent on media” to drive their attention to the right consumers. 

In your own words, tell us why a brand marketer or publisher would want to use Lotame’s Audience Modeling solutions.  

If you want to use data to power campaigns on the open web, there’s really no one else that is sitting on as much third-party data as we are, and that is able to take that data, export it and make it usable in a number of platforms. Sophisticated modeling plus activation is really the Holy Grail for anyone interested in customer acquisition. With walled gardens, you’ll only live in their world. 

So there you have it. From the “math man” himself, Omar Abdala shares a sneak peek into Lotame’s Audience Optimizer and how it can help brands extract more from their customer data, find high-value audiences, create winning prospecting strategies, and drive performance across screens. 

Ready to learn more? Read our Audience Optimizer FAQ’s here.