An Interview with Professor Joseph Konstan on the upcoming MOOC, Introduction to Recommender Systems
One example [of recommender systems] is with networks of websites where, the first time you go to a site, it is personalized to you based on things it learned from other sites in the network. It shouldn’t be surprising to someone if they are looking for an expensive car, and after browsing for a while, … Continued
One example [of recommender systems] is with networks of websites where, the first time you go to a site, it is personalized to you based on things it learned from other sites in the network. It shouldn’t be surprising to someone if they are looking for an expensive car, and after browsing for a while, they see an ad pop up with the same car brand that they were looking at four websites ago.
Recommender systems are playing an increasingly important role in providing personalized content to users online. Class Central sat down to talk with Professor Joseph Konstan, of the University of Minnesota, who, along with Michael Ekstrand, is teaching a MOOC on recommender systems via the Coursera platform. The course is free to enroll online, starts on September 3, 2013, and lasts 14 weeks. An edited excerpt of the interview with Professor Konstan is below, moderated by guest contributor, Charlie Chung.
Can you tell us a little about your background and how you became interested in recommender systems?
I got involved in recommender systems in 1995, when my colleague, the late John Riedl had published the first paper on collaborative filtering, which was one of the early techniques in recommender systems. We started working together to grow that concept, first into a larger research prototype that we tested, and later into a spin-off company and a larger research agenda that has grown. I was there for the early workshops and founded the first conference. I’m still working in the field of recommender systems today, but also in the broader area of social computing systems that use inputs from lots of different people to deliver value to each of them.
Can you tell us a little about the history of recommender systems and what technologies or other breakthroughs enabled them to develop into their current form?
Recommender systems depend on the ability to gather information on peoples’ opinions on content, and it blossomed in the early days of the world wide web, though the early systems were not web-based. It was the web that provided the interactive environment where people could be asked to rate content and it also enabled implicit measures: how long people looked at something, what they clicked on, what they bought, etc. In the 1990’s, commercial interest further fueled the evolution of recommender systems, where the large scale involved required creativity and a rapid evolution in computational techniques. Another milestone along the way was the ‘Netflix Prize’, which offered a million dollar prize, and got a lot of people from fields that had not been involved in recommendation to try to solve problems in recommender systems.
Are there any surprising uses of recommender systems that the average internet user might not be aware of?
I think you’re going to find that there are recommender systems almost everywhere these days, utilizing a variety of techniques. One example is with networks of websites where, the first time you go to a site, it is personalized to you based on things it learned from other sites in the network. It shouldn’t be surprising to someone if they are looking for an expensive car, and after browsing for a while, they see an ad pop up with the same car brand that they were looking at four websites ago.
In the early days, recommendations were mostly explicit, because companies spent a lot of money on their recommender systems and they wanted to be in the customer’s face saying “Hey look, we are helping you. Here are suggestions!” Now, I think more sites simply have the technology in the background selecting things for you. After all, if companies are putting products on a screen, they are probably going to pick them based on some form of recommendation, and it would be silly to say every time, “Oh, by the way, this is put here because we think you’re going to like it.”
What are some of the latest trends and leading-edge work in recommender systems that’s being done right now?
There is a lot of work in recommender systems in the technology of how we compute more efficiently, lower the overhead, and be more dynamic. We’ll talk a little bit about that in the course. But the work that I find the most fun involves interesting user experiences, such as context-aware applications. For instance: how do you adapt a recommender if we know your location via your mobile phone? If we’re suggesting a meal to you and we know you’re at the beach, can we suggest something appropriate nearby? How can we adapt recommendations based on who you’re with, or the weather, or even your mood, if we have that information? Another area where a lot of interesting work is going on is in interfaces to recommender systems that give the users a greater degree of control and visibility to enable better decision making.
One application that is growing rapidly, that I think will be really important, is education. The ‘holy grail’ concept is this: as we are watching someone go through a set of online lessons, we figure out, based on what they’ve learned and what seems to be working for them, the best path for them to master the next set of material. In the past, there was never a sufficient size of datasets available to work with. But as we’re getting more online education materials, this is suddenly becoming possible. And I think the idea of having a meta-learning tool offering guidance is a really exciting concept.
Regarding the class, there is a Concepts track and a Programming Track. What types of students is each track intended for and what should they expect to learn?
The Concepts track is for people who are interested in learning all about recommender systems but are not interested in learning how to program them. Our goal was to meet the needs of three different types of people: 1) Those who could do the programming work, but don’t have the time to commit to do it, 2) Those who may not have the programming skills yet to do all of the assignments (our Programming track requires some serious Java skills), and 3) Those who are not interested in programming, perhaps business students or marketing professionals, but do have gone through college math, to understand matrices, manipulations, etc. I think all three groups would find the Concepts track to be useful, because we will go through a lot details of how recommendations are done and cover important topics, such as trust-based recommendations, the impact of recommender interfaces on user behavior, etc.
The Programming Track is for our campus students who are enrolled in this course and for anybody who wants to master the programming of recommender systems. It includes everything in the Concepts Track, plus six programming assignments, five of which will be based on the LensKit recommender toolkit. We’ll actually have students implementing recommenders, tuning them, and running evaluations, all in their own Java environment.
If technical programmers are interested in getting into this field, how much would going through this course prepare them to work on actual recommender systems out there?
Pretty far along. LensKit, which is open-source and available, is not designed to be commercial-scale, but it has all of the same interfaces as a commercial-scale tool. Thus, for someone who wants to go out and work on recommenders in industry, this course would probably be as good a preparation as anything that is available.
How many online students do you have enrolled at this point?
Close to 20,000.
How do you feel about that?
On the one hand, I think it’s just really exciting. It’s humbling for that many people to want to take a course in a topic that is still a niche specialty. On the other hand, it’s a little bit scary, because anything that we don’t get right is going to affect a lot of people. But my co-instructor, Michael Ekstrand and I are well-prepared, with carefully designed assignments, and a TA who will monitor the discussion forums and keep an eye out for significant issues. We will use a dedicated recording studio to record the lectures. We want to make sure we give everyone the best experience we can.
Also, this will also be a flipped classroom for all of the students in our University of Minnesota campus class, and they will be getting their lectures through this MOOC. We plan to record 8-10 of the campus discussion sessions and post them for viewing for Coursera students who may find them helpful.
I think this is going to be an exciting experience and I’m hoping that we provide something that many of the students around the world would not get a chance to have otherwise.