Data Science & Artificial Intelligence In Digital Publishing
It’s been said that marketers ruin everything. I certainly wouldn’t go that far, but am willing to admit that modern marketers can certainly make it difficult to fully understand nuanced subjects on sophisticated topics like artificial intelligence and data science. Digital publishers are now starting to see just how muddy these waters are getting…
My good friend John Cole was on a recent AdMonsters panel with Google discussing the growth of data science and emergence of artificial intelligence in the digital publishing space. The conversation got quite interesting when they really started to define what these terms actually mean (vs. what marketers will tell you they mean).
Below, I’ll share a couple of things that came up during that panel, as they were interesting and bring to light some of the issues I think a lot of digital publishers are struggling with. I’ll also highlight some examples of how these technologies and concepts can be applied to modern digital publishing operations.
What is artificial intelligence and machine learning?
I’ll try to keep this at a high-level and in-context of the world of digital publishing; otherwise, I’m afraid this subject could get convoluted pretty quickly.
Artificial Intelligence is a subset of machine learning. Machine learning is the overarching concept of a computer specifically taking a type of data and learning from this particular data over time (in many cases so that decisions can be automated). Artificial intelligence is a sub-vertical that takes this data and “learns” in the way a human brain does; mimicking the behavior of neurons and the alike.
So how does this apply to digital publishing? Well, remember those pesky marketers I talked about earlier? One of the really interesting things brought up in the AdMonsters panel was the abundance of technology and services being marketed to digital publishers right now that use the terms “A.I.” or “machine learning” that actually are not correct applications of these terms.
In many cases, marketers are choosing to label algorithmic-driven technologies as “A.I.” or “machine learning”. Although the difference may seem slight, they are actually very different.
What’s the difference to digital publishers?
In many cases, these terms are being convoluted most when discussing things like ad targeting, ad testing, and various other forms of ad operations and data science. Identifying where machine learning and A.I. actually exist in these spaces could make big difference when publishers are defining future strategies.
The principal difference between machine learning (something like automated multivariate testing) & algorithmic optimization (A vs B testing) is this: Algorithmic optimization requires a human to not only decide the factors included in the calculation but also hard-code a set of if-then rules that determine how decisions are made. They are locked.
In other words, algorithmic optimization means that the rules are written for the machine, and with machine learning, it’s the reverse — the machine writes the rules.
Putting artificial intelligence into practice
Let’s say we want to know if CTR is connected to geo-location (so that we could potentially adjust our bid floors by geo). Using algorithmic optimization, we’re looking at just a few factors — the line item, the geo, the historical CTR, device, the upstream traffic source and so on.
We might write a program to statistically predict the probable value of an ad over time (from CTR/CPC/eCPM) and plot revenue improvements from the adjusted floor changes; but the major flaw in this sort of calculation is that there are likely to be missing factors that are critical to the overall objective (in this case revenue), but because our (written by us) algorithm is restricted to what we’ve decided is important.
Algorithmic-driven approaches to this strategy would miss a lot of valuable opportunities. For example, the dilutive effect of other ads on the page – their own price profiles and their effect on subsequent pages / subsequent ad values (i.e. – should that ad on tablet devices even be there for that user, if we want to raise things like engaged session duration and revenue as an outcome?)
Machine Learning takes all the signals given by a user and weights them for the publisher by training it’s own algorithms/calculations which then adapt over time. This is the main difference between the kinds of compounded results you can get from A.I. vs algorithmic driven solutions.
This is ultimately a more sophisticated and more personalized approach to solving these challenges. It all may seem the same to some publishers, but as the market evolves at a breakneck pace, there is an increased call for tailored user experiences, and this is something only A.I. can actually provide.
An industry data scientist recently explained this at a MediaTel event along with The Guardian, Financial Times, and Hearst Media. You can watch a highlight of that video below.
Why testing may be the greatest application of A.I. for publishers
Mark Zuckerberg announced that at any one time, there are 10,000 versions of Facebook live at any one time. Here’s the full quote:
“At any given point in time, there isn’t just one version of Facebook running, there are probably 10,000. Any engineer at the company can basically decide that they want to test something. There are some rules on sensitive things, but in general, an engineer can test something… And then, they get a readout of how that affected all of the different metrics, and things that we care about. How were people connecting? How were people sharing? Do people have more friends in this version? Of course, business metrics, like how does this cost the efficiency of running the service, how much revenue are we making?” – Reid Hoffman interview with Mark Zuckerberg – ‘Perfect is Imperfect’ – Masters of Scale Podcast
I think this is a good indication that the testing ethos is expanding. Google also said on the AdMonsters panel that their search algorithm (for search results) is now largely driven by A.I. and not individual updates (e.g. the days of one big update e.g. Panda — are over).
This quote also brings to light something obvious. Most digital publishers, even the smartest ones, aren’t doing the kind of testing that platforms like Facebook and Google are doing. This is one of the biggest untapped opportunities for most publishers; as testing that affects UX has more potential than just about any other strategy to increase user experiences and revenue.
Where does data science fit into all of this?
A definition of data science I heard recently from a digital publisher was the best I’ve heard. They said that data science is a combination of statistics, scientific method, and computer science — a mix able to make intelligent predictions using data to maximize the desired outcome.
The evolution of digital publishing will see more companies embracing data science because of it’s obvious connection to technologies like true A.I. I think this will mean more personalized experiences for users, more testing, and more data-driven decisions that don’t rely on things like surveys, user feedback, and personal preferences (as publishers know these things can be hit and miss).
Ultimately, these changes for digital publishers can have an intimidating — and frankly insincere — feel to them; however, being exposed to what these technologies can really do and the problems they can solve will ultimately prove to be empowering to digital publishers.
What do you think?