Catherine Williams, Chief Data Scientist at AppNexus, speaks to ExchangeWire about the launch of APB and how it is helping marketers find “a third way” to optimise their ad spend through their use of their data.

EW: Some people have described the launch of your latest APB, as a ‘bidder in a box’. Could you explain a bit more, and what kind of problems it will help solve?

CW: The primary problem that this is solving is that you have two extremes if you are a marketer who wants to use data to optimise advertising spend.

On our platform, you can use our built-in algorithms that use 10 variables, and you’ll okay outcomes, but with that, you’re not using the full potential of your data assets.

There are other parties you can hand your data to, but that’s a black box – we like to think we’re very transparent – and they will take all your data and say: "We will optimise all of this for you, but you don’t get to retain control of your data."

So, that’s one end of the spectrum, basically you hand over your data and get ‘whatever’ results.

On the other end of the spectrum, you can build your own bidder, which is very doable as there are open standards.

However, that’s an enormous amount of work, even though you might think that it sounds very simple in practice.

Then you realise the enormous amount of work you have to put in, and all the different use-cases you have to support, and the latency, throughput, scale, and it becomes an enormous amount of investment before you can start using it to make money.

So, you find that a lot of companies that go off and try to build bidders end up coming back. While they may have an enormous amount of data that they want to make use of, they don’t have a good way to do it.

APB is like the third-way that lets you build models with whatever amount of expertise you have in-house that lets you extract value from the data you know it has.

Then you can just plug it into our bidder and it can just go, without having to worry about all the fragility – or the huge engineering teams the other method would need to stand it up – using our scale, on our inventory, with all of the support we can offer.

EW: Previously, this wasn’t possible, what’s different now?

CW: We don’t want people to give us compiled pieces of code that we execute, as we don’t know what it will do; as we haven’t tested the code, and won’t know if it will take down a machine.

So, while there’s a difference in executing a code, and executing an application, which is not something that we are allowing now, or ever (most likely).

What this will allow, however, is to build a model; and that’s the key thing that most advertisers want anyway. When you build an application, it is to run a model that will output a prediction or a bid-price. What we allow is that you put in just the part that’s novel to you, and make everything else a commodity.

EW: Some people would say that this would rival the offerings of IPONWEB, Datacratic, and Shiftforward. What makes the APB offering unique?

CW: I can’t speak for those other companies, but from my understanding, some of the companies you listed above help you set up your own bidder (I believe), but I don’t think they give you an easily ‘pluggable’ way just to run your own models.

For instance, if you had just a data science team, then you could use APB, but I think with those other companies you would also need to have an engineering team, and a data science team.

EW: Having spoken to some of those attending your Optimize summit, we wanted some points of clarification after the early presentation of APB. Does this mean the advertiser has to write the code themselves before being plugged into APB?

CW: The way our data science team did it was to use an off-the-shelf, open-source algorithm to build the [decisioning] tree, so they didn’t write the algorithm to generate the tree. They may have performed some manual tweaks, and some scripting, but they didn’t write it.

This is then outputted into whatever to the relevant native format is of that package. However, the part that they do have to do is taking the output of that algorithm, and then translate it into Bonzai [one of the key features of APB], and that was pretty simple. It was just a matter of formatting.

We have tried to build in support for major open-source algorithms to make it easier, but that’s the only coding required today. Advertisers can always generate their own algorithms to generate trees, or whatever they choose.

EW: So is APB only realistically open to advertisers that have their own team of data scientists?

CW: I think that’s only one use-case; advertisers that have data scientists, but without a meaningful way to plug their insights into their advertising spend.

Although, I think there’s other cases as well. If you look at some of the more innovative programmatic media companies on our platform, they are hacking our system in complicated ways and building hundreds of campaigns, with complex optimisation.

I don’t know who’s staffing that, or if they’re calling themselves data scientists. I think they’d just call themselves traders. These are savvy people who look at reports, and figure out a way into the system.

Those types of people can also use this. These people can build one tree worth of logic just by hand (which is much easier than creating hundreds of campaigns), plug that in, and it works just as well.

So, I think there’s a spectrum of people who’ll really benefit from it.