Social Choice for Fair Recommendations

**Kyle Polich** (0:06)
Welcome to Data Skeptic, a podcast exploring the methods, use cases and consequences of recommender systems.
Welcome to another installment of Data Skeptic Recommender Systems. Today on the show, the centerpiece of our discussion, which we'll eventually get into, is called SCRUF-D, that's an acronym for social choice, for recommendations under fairness, and then we throw on the D for dynamics, SCRUF-D. This is a novel, agentic approach. As we get into the everything agentic world, the core concept is take every fairness concern you might have and represent it as an agent. Each agent then, each iteration acts in the best interest of whatever they're trying to accomplish. That's the allocation step. Then there's a preference aggregation step in which is something like voting happens, where these agents all come together and decide on what recommendation to give. We'll get into all that and more during the interview.
I'm Robin Burke. I'm a professor in the Department of Information Science at the University of Colorado Boulder. I teach a wide variety of classes in information science. I teach machine learning. I teach network science. I teach other data science topics, recommender systems, courses, and other sorts of data analysis sorts of things.
When did recommender systems first get on your radar? I got my PhD in 1993, so I've been around in this academic scene for a while. And if you looked at that dissertation today, which was really about recommendation in the context of intelligent tutoring systems, we would call it a recommender system, but that term did not exist in 1993 So, in some ways, you could say, like the path of my dissertation research was sort of a recommender systems dissertation. But again, that term didn't exist. So then you would have to kind of go back to maybe like 1989, 1988 to say, okay, when did you actually start working on recommender systems? So that's one kind of way to answer that question. Another, I think, important connection is the connection to case-based reasoning. So case-based reasoning is a kind of AI technology that was contrasted to say, rule-based reasoning or expert systems that were very prevalent at the time. To say the way that people solve problems is by reference to things that they have solved before. And so the question is not, how do I reason about this from scratch? The question is, how do I find something in my memory that's appropriate and adapt it so that it meets the needs of the current problem? If you just grab on to the retrieval piece of case-based reasoning, it looks a lot like what you would expect of a recommender system. And so there is a close connection there, which I think doesn't get recognized that often, but that's the route that I came into recommender systems. So once I finished my PhD, I went on and started a postdoc at the University of Chicago, and the professor I was working with there, Professor Christian Hammond, who is now at Northwestern University, he was trying to apply some of these same case-based reasoning ideas to things that were consumer type domains. And he was very much interested in this, like, okay, when the solution to somebody's problem is a product, like a new car, how do you find the appropriate item?
That was the sort of beginning of what was, at that time, we called knowledge-based recommender systems.
And that's how I started working in that field. And again, term didn't exist for a couple of years after that. And then once people started talking about recommender systems, the other folks who were working in the area, that term came into being. We realized, oh, that's what we're doing. And so that became the kind of umbrella term for all of these kinds of technologies.
As I look through your long list of publications, one that jumps out, just because it's incredibly well-cited, is a Hybrid Recommender Systems Survey and Experimentation. So just a little while later, that word's there in the title, We Have Recommender Systems. As a point in time, if you can recall that paper, what was the state of the art? How did people approach recommender systems? Nowadays, if people are building AI systems, you basically put everything in it. You throw in all the features you have, the user data, the item data, meta data, anything you can get your hands on that might help you sort of unpack what the user-item interaction was, you try and make use of it. But this was not true at the time that I wrote that paper. And at that time, there was kind of these two camps. Well, it was like two and a half camps.
So one camp was the people who were doing content-based recommendation, and they were saying, look, you want to understand the features of the items, you want to learn what it is that the user attends to when they're thinking about, say, a movie, the genre of the movie, the actors, or other kinds of consumer items. What is it the user is paying attention to in terms of the features of the item itself? And then the other camp, there were the collaborative recommendation people who are saying, no, no, no, feature catalogs are unreliable, hard to construct. All you really need is other users, their patterns of interest. You can uncover the important features. That's the whole logic behind collaborative recommendation, is that maybe these content features are not as necessary because all of that is recoverable from the signal that you get from other people's preferences if you have a big enough user base. And then I say two and a half camps because the other camp was the one that I was started out in, which was the knowledge based camp that basically said, in addition to the knowledge features and in addition to this sort of user preferences and things like that, it's useful to have domain knowledge and to know that the relationships between different kinds of features. If you are going to have a faster CPU, it's probably going to cost you more and you're going to need a bigger power supply, things like that, like relationships between things. Those three kinds of recommendation coexisted at that time. It was a back and forth as people saying, this one is better for that reason, this one's better for that reason.