Sramana: Recommendation engines are very difficult to do. How did you build your algorithm?
Tobias Bauckhage: First, we filtered by domain. Users may share similar interest in comedies but have completely different interest in horror films. It is also important to understand the content. You see people doing collaborative filtering or a deeper “DNA” analysis, which is how Pandora analyzes music. If we had a million users, then we can group them in the thousands who have similar tastes. Then we can look at the categories, such as car chase films, and see what feedback has been within that group and make recommendations to others in that group who have not seen those films.
There is another dimension that has entered the scene, which is the social dimension. In that case people will watch the movie even if they don’t like it just because they feel that they will be left behind socially if they don’t. Finally, location has emerged as an important element. What is the size of your screen? Is it Monday after work or is it Saturday? A new generation recommendation engine should be able to understand where I am right now and factor that into the recommendations produced.
Sramana: This is very interesting. I have done a lot of work in AI and machine learning. This is contemporary and very much needed by people who are building systems. This is a modern day problem that people are trying to solve in sophisticated ways. I’ve found some good things through Amazon, and I think Netflix recommendations are fairly good as well.
Tobias Bauckhage: Netflix has some interesting challenges now. You don’t really need recommendations there because their catalogue is so limited now. In 2009 we were at a point where we were getting very sophisticated with our recommendation engine. We loved the intellectual challenge, but we could see that the business model was limited. We still had a huge community growing, but there were just not that many takers for the software. We would only have 40 potential clients, and the B2B industry has a long sales cycle.
So, in 2010 we had to face the reality that we had put a lot of effort into an interesting problem, how to find the right film for any fan, but it wasn’t all that important a problem to solve from a business perspective. It was a luxury problem.
At that point, we had a third co-founder coming into the company. We knew that all the work we had put into solving the problem created interesting algorithms. We could find the right film for our fans, so we just turned the algorithm around and set it to finding the right fans for films. That is a much bigger problem in this industry. That addresses the crucial 48-hour window. You have to mobilize your audience to see a film the first weekend. That means you must know your audience and the reality is that nobody truly did.
We wanted to know who was following a film six months before it came out. Who watched the trailer more than once? We had the right knowledge about taste and profiles of users. We were able to go back and build products around that for distributors and movie companies. Of course, we would never do that for an individual profile, just a community of users. From there we had to invest in our advertising capabilities. We still had a B2B business, but we did not have to invest much more into the business. We were able to concentrate on the critical mass that we had built.