At IBC, Android TV PM Discusses Mixing Machine Learning and Content

At the International Broadcasters Convention in Amsterdam, which was held recently, Android TV PM Sascha Prüter talked about the recent FCC rules that would allow cable providers to make their content available through an app. This convention, held annually, had a number of discussions and announcements including one from Technicolor.

Sascha talked about the problem with the app approach. “Apps are not the ideal way to transport content. They create a silo and once you leave that silo to watch other content, you are in another silo.” The siloed approach has been something that the Android TV has tried to tackle in a number of ways. The system has a universal search feature, which allows any content provider to expose their metadata so a user’s search will search across everything. There’s the recommendations row, which allows any app to suggest content for the user to watch right on the homescreen. There is also Live Channels, which allows any app to create channels of streaming, linear content.

He also discusses ways to alleviate the silo problem by applying machine learning. Machine learning has been an area of growing interest for Google. They’ve even manufactured custom computer chips for running machine learning processes. Machine learning could also be applied to content discovery.

A content assistant, which knows about the video content of each app, would be able to recommend content if “a user was tuning out of a show 15 seconds after queuing it up”. Then that show would not be rated highly again. Measuring engagement would be an easy metric to use. Prüter suggested “a scenario where people could wear sensors that provide information allowing their TV to predict what kind of content they are in the mood for.” This could easily take the form of Android Wear. If it measured your heartbeat, it could tell how invested you were with a particular video. If you combined the two, it could tell you if that horror movie actually did scare you or if that action movie had you on the edge of your seat.

It could also tell if you were bored. Are you shifting around a lot? That may indicate you’re not watching. Are you not moving at all? That may indicate you’re asleep. Netflix created socks that pause the video when you fall asleep. Even interacting with the notifications on your watch mean you’re not engaged with the current program.

There are other metrics based on your personal data. Was your calendar full of appointments? You may be stressed and want to watch a comedy. Is it a rainy day in your town? Perhaps you want to watch something long and dramatic. Have a date with someone else on your calendar? Perhaps it can use both of your preferences to find the perfect movie. When you start to think of a recommendation engine, there are a lot of different metrics you could use.

However, this assumes that the content is being treated equally. It would be easy for the engine to suggest imperfect results based on hidden biases. “[I]nternet companies are more likely to experiment with pointing to free, available content first while cable operators are more likely to want their content to come up first.” Recommendation engines need to be good at their job in order to get users coming back to it, and keeping user trust is perhaps most important when using a lot of their data.

Nick Felker

Nick Felker

Nick Felker is a student Electrical & Computer Engineering student at Rowan University (C/O 2017) and the student IEEE webmaster. When he's not studying, he is a software developer for the web and Android (Felker Tech). He has several open source projects on GitHub ( Devices: Moto G-2013 Moto G-2015, Moto 360, Google ADT-1, Nexus 7-2013 (x2), Lenovo Laptop, Custom Desktop. Although he was an intern at Google, the content of this blog is entirely independent and his own thoughts.

More Posts - Website

Follow Me:
TwitterLinkedInGoogle PlusReddit