The data paradox of a predictive TV recommendations engine

TiVo's next-gen user experience technology on display at IBC 2016.
Ben Munson

AMSTERDAM – At this year’s IBC conference, a strong theme revolved around making recommendations more predictive using contextual, time-based data. Paul Stathacopoulos, vice president of strategy at TiVo, said the goal was to make the predictions so good, that viewers didn’t even need to go into the guide anymore.

Indeed, TiVo has put its new predictions feature front and center on its newly redesigned user experience and backed it with loads of data on what people watch, when they watch, and on what device.

The push to improve recommendations across pay-TV and VOD platforms is not new. It’s already been 10 years since Netflix first offered a $1 million prize for an algorithm to improve its recommendations. And major MVPDs have been pushing to improve the search and discovery on their set-tops, like Comcast and its major investment in its X1 platform.

But if a predictive recommendations engine that runs on viewer data becomes so good that the viewer no longer has to make decisions and therefore isn’t generating new data, how can the engine continue to make predictions?

TiVo was hardly alone in advancing this admittedly hypothetical paradox. Companies including Alticast and ContentWise were talking up a new platform partnership to drive more personalized TV through the use of tools including predictive browsing and context awareness. At a Sunday panel discussion on the future of UX and the evolving TV experience, participants like Android TV head Sascha Prueter predicted that TV platforms will soon feature “content assistants” that can use machine learning to better understand viewer behavior and provide a simpler way to select content.

Likewise, Steve Allison, senior evangelist for Adobe Primetime, was at the show to discuss how Primetime’s new ad replacement technology is powered by a recommendations engine that uses the contextual awareness of the user to help refine the recommendations is makes.

“Primetime, since its beginning, has been collecting consumption data about what people use. So now we’re putting all that back together again and that enables us to say, at this particular point of time, this person or this audience segment tends to watch this kind of content,” said Allison.

So, I asked him, if viewers start to solely rely on the predictions instead of going into the guide, where do we get the data that is used to make the predictions?

“That’s an interesting one,” replied Allison. “It doesn’t really matter if they don’t go into the guide because of the granularity of the data we’re collecting. Even if they don’t go back, once we start identifying things in the client and in real-time, then we can start pushing content toward them.”

“But it’s only going to work for certain kinds of video consumption. People still want to browse through what they’re going to watch on Friday night with their mates. The kind of content that you’re using will have a big impact on how you do the recommendations.”

Even TiVo, as it searches for ways to reduce the time viewers spend sifting through channels, looking for something to watch, understands that viewers will still browse.

Michael Hawkey, SVP and general manager of the discovery business group at TiVo, admitted during an IBC panel that surfing channels will always be a thing and that, honestly, a lot of viewers just find it relaxing.

For now, then, it seems like the industry is confident that recommendations will continue to get smarter and more predictive while consumers will continue to the boldly quest for new content. That means that the solution to the discovery problem that providers and programmers face will fall somewhere between calculating the next step and going with what got us here. -- Ben