Meet the Intern Who Knows What Videos You Want to Watch

I saw that that kind of information could be useful for our purpose, our prediction, our recommendation. The algorithm we used couldn’t incorporate all that information. I did some research and found an algorithm called Factorization Machines which could be used to do that. I made some modifications to the algorithm and we tested the data. The results are really good. Using the context versus not using the context, the best case we have shows a 20 percent improvement in prediction accuracy. Geo information, device information, and also some feature information about the video: the content information. We also use information about the session: At what time did the session happen and on which website? We refer to all of this information as context information.

Streaming Media: Tell us more about the information that was being discarded.

Gang Wu: Basically, there are three groups of information. The first group is feature information about the user: the user’s gender, age, and education. The second group is about the content, like the language, the category, and the duration. The third group is session information: What time did this session happen? Morning, noon, or in the evening? Also the location that this session happened.

In the previous recommendation system, we only used the viewing history. This kind of site information or counter-information was discarded by the algorithm.

The source of the data doesn’t matter. In our experiment, we dealt with data from different sources. The data we use is from Adobe Analytics. Adobe Analytics can collect the information we want from different devices: browsers, TV, Xbox, PS3.

Streaming Media: How did you test your algorithm improvement?

Gang Wu: Its accuracy has been tested with Adobe TV [Adobe’s in-house video system], and also in Primetime.

After we got the algorithm ready, we tested it with an Adobe Marketing Cloud customer. We saw good results from that experiment. The Primetime product manager showed the results to a customer, and the customer was very satisfied with the result and showed strong interest.

Because of that positive feedback, the Primetime team decided to implement this recommendation feature into Adobe Primetime. It happened last fall. The engineering team started to implement the prototype last fall and around this spring, it has now finished implementation. We want to see whether the engagement of the user has been improved. If there’s a strong change then we’re going to show that to more customers to see whether they’re interested.

Streaming Media: What does it mean that Primetime’s predictions are more accurate?

Gang Wu: Because of this, customers can make better predictions. They can predict users’ potential interests better.

Streaming Media: How else can you improve the recommendation system?

Gang Wu: That’s what we’re working on right now. Adobe Analytics can collect 1,000 pieces of information for each session, but we cannot send all that information to our algorithm because a lot of it is not useful. What we’re doing right now is manually taking some items that we think are useful. In the future, we want to make our algorithm capable of automatically picking the information that gets used. We call that automatic feature selection.

Streaming Media: You’ve spent a lot of time at Adobe. Do you think you’d like to make a career there?

Gang Wu: Yes, I think that the feedback for the work from our team at Adobe and also from our customers is very positive. Also the company has a strong track record. It’s very likely I will join.

After the interview, we asked Vishy Swaminathan how the discarded information resulted in improved recommendations. Here’s what he said:

We believe contextual information about user sessions plays a major role in what videos should be recommended. Context refers to time-of-day, device, location, geography, or anything else that affects the users’ choice of video.

For instance, what someone wants to watch in the evening on Apple TV is very different from what that person wants to watch on her mobile phone waiting in line for coffee or in a train. Same user, but different context.

This rich context information is something we collect in Adobe Analytics for every user session, but before Primetime Video Recommendations it was never combined with user consumption for recommendation. This is used as context that feeds Adobe’s algorithm for recommendation, coupled with existing information from session completion or ratings data, user data, and video metadata.

[This article appears in the October 2016 issue of Streaming Media magazine as "Meet the Intern Who Knows What You Want to Watch."]