Meet the Intern Who Knows What Videos You Want to Watch
Great ideas can come from anywhere. Even an intern can come up with a brilliant idea that changes how millions of viewers watch over-the-top (OTT) television.
Gang Wu is a graduate student at Iowa State University, where he studies matrix completion (more on that later). For the past 4 years, he’s spent his summers interning with Adobe at its headquarters in San Jose, Calif. Wu is part of a team analyzing video recommendations, and earlier this year he came up with an idea for bettering Adobe’s recommendation system. His algorithm improvement went through the proposal process and ultimately became a part of Primetime. It shows that every member of a team can contribute, and a smart company listens to all its employees—even the interns. We’ll let Wu tell the rest.
Streaming Media: Hi Gang. Start us out by telling a little about yourself.
Gang Wu: I was born in China. I got my bachelor’s from China at Dalian University of Technology, and then I came to the United States for my Ph.D. at Iowa State University. From my second year of my Ph.D., I got an opportunity to work at Adobe Research as an intern. Viswanathan (Vishy) Swaminathan [principal scientist for Adobe Research] is my advisor. My Ph.D. is focused on matrix completion, so it’s actually related to the problem Vishy was facing, the recommendation problem.
Streaming Media: What is matrix completion?
Gang Wu: When a company like Netflix has a matrix with some user ratings, they want to predict some of the missing ratings. That’s very useful for recommendations because if you can predict the missing ratings, you can know people’s potential interest on movies that they didn’t watch. Imagine that there’s a matrix, and that each row is a user, and each column is a video, and the entries are the ratings. Some ratings aren’t known because some people haven’t watched some of the movies, and most of the entries are missing, because there are a lot of movies that haven’t been watched by the users. Companies want to predict those missing ratings, so that’s matrix completion.
Streaming Media: How did you get involved with Adobe?
Gang Wu: My advisor at the university had another student previously who also worked as an intern at Adobe Research. He graduated and he got another position. Vishy talked to my advisor at the university and asked whether there’s a student that’s really interested in their problem. My advisor talked to me, and I was really interested in the problem, so I came to Adobe for the internship. That was in September 2012. I’ve worked with Adobe on and off since then as I finish my Ph.D.
Streaming Media: What department are you working in?
Gang Wu: I work in the Systems Technology Lab, and we have researchers focused on matrix theory. They have a lot of good publications. Also, we have researchers that work closely with the product. Those researchers are very familiar with the product, and they work on prototypes for new technologies or products.
Streaming Media: You’re working on a matrix completion problem called the Netflix Challenge. What is that?
Gang Wu: The Netflix Challenge is about user ratings. The user needs to give some manual input to create the ratings. In the real world, most users don’t give any manual input. They don’t give ratings after they finish a video. That’s an issue. If we don’t have ratings, how can we still make good predictions and recommendations? Vishy came up with the idea that we don’t need users’ manual feedback, because the product can track some user activities, like the user’s clicks and the user’s keyboard. The product can also record how much of the video has been watched by the user. Vishy came up with the idea that we can use that automatically recorded data to create predictions or make recommendations for new videos for you to watch. That’s the problem I faced when I initially came to Adobe in 2012.
After my first internship session, we came up with the idea that we can record session progress to replace the manual rating. For example, we can look at the duration of the video that was watched by the user. If the video is 10 minutes and the user watches 5 minutes of that, the session progress for that case is 50 percent.
Streaming Media: What was the innovation that you brought to the product?
Gang Wu: I’m the person who did the cleaning and structuring of the data. It was very messy; there was too much information there. The actual data we feed into the algorithm to make predictions is very compact. I came up with the idea that the information we throw away when we do the cleaning might be useful. For example, in the data I deal with, we have information about the user’s device, the content, the language of the video, and so on.