Welcome back. Today's interview we've brought back Tony Jamison Principal Researcher at DFKI German AI labs. To talk about an interesting challenge that we don't talk about nearly enough in recommended systems. How do you recommend not for an individual, but for an entire group of individuals, each of whom may have their individual preferences? Welcome back Tony. >> Glad to be here Joe. >> So why do we start with just an example of the kind of place where group recommendation has shown up in the past? >> Yeah, we can take one of the very first examples from the late 90s. This MusicFX system was designed for a gym. So imagine you often go to a gym, and while you were working out, music is been played through a loud speaker so everybody hears it. So then the question arises sort of music should be played? Now the MusicFX was in a studio where they had 91 radio stations, each of which had different styles of music. And consequently, the problem was to choose 1 of those 91 radio stations which was a good match for the musical tastes of the people who happened to be there. Who may have been 2 or 3 people, or maybe up to 25 people. >> So obviously if I were doing this I know what music I would like, but how do you pull together music for 3 or 5 or 25 different people? >> Well, first thing you have to do is figure out somehow what their musical taste are and their ways to doing this, what they did at MusicFX is quite simple. They asked people to fill in a form, like the one you see here. These are three scales with five points ranging from hated to love it, and each person would have to go through and check one of these boxes for each of these 91 radio stations. And this was the basis they use for it using stations for the group. >> Interesting, well, we'll come back to this example, but can you think of other examples where we need to make a selection that matches a larger group of people? >> Well one popular system these days is Doodle which is used for arranging meetings. And here as some of you probably know what each participant does is go through sort of calendar and indicate when they're available and when they are not, and sometimes in intermediate state. So here this fourth person who would be just about to introduce her preferences for times. >> Neat, so I think what this raises is that, we have some interesting choices, as to how we make decisions. And how we make the recommendations given all these different inputs, and since we have Doodle on the screen, why don't we start there? And we can see right now it looks easy because there's one time everybody can make it, but when it doesn't work out that easily. How do we end up with a meeting time from Doodle? >> Well with Doodle it's pretty easy because it's not actually a recommender system it just collects the information, and whoever's organizing the meeting goes in and makes the decision. So, that person is essentially doing what the group recommender would have to do, and then they figure out what the best solution is. That means that they can be pretty clever for example, that David is the big important boss and Richard is just a lower level person. Now the group leader may want to give the President's to David's taste. >> And Doodle gives a little bit of help by showing this tallies on the bottom, as to how many people can make it to a given time. And I know from my own experience one of the things I've always done with filling out Doodle balls, is if nobody could make it time. I didn't even bother marking it anymore because, I knew nobody would pick a time that big boss David wouldn't make it. So why bother with typing an extra clicks? >> Exactly that's a good example of why the so called specification of preferences is quite different for group recommenders and for individual recommenders. That if you happen to be one of the later people to fill in your times. You can save time simply by checking what other people have marked and just simply not bothering with the dates like this Tuesday the 13th which are obviously not going to be chosen anyway. So in that sense preference specification for group recommenders often takes quite different forms than it does for individual recommenders. >> Well let's go back now to the example of MusicFX. You obviously can't have somebody sitting in the gym reading through all of these profiles and picking a station. You need something that's fairly automatic because what we want is for people to focus on exercising. So, how do MusicFX decide of which genres to play? >> Yeah, essentially what it did was give an overall rating. >> So how did MusicFX do this? >> Essentially, for each radio station, you tried to figure out the overall rating of that radio station by the group. Let's look at this example, here we have four other people in the gym besides you, Ann, Betsy, Chuck, and Dave. And they have given these rating here. So intuitively, you would think that the overall rating for the group at this point would be this fourth step here. Now to make it a bit more precise, is suppose you're giving your own rating. Let's say you actually happen to like step four. One thing the system could do although it's not exactly what it did it is compute the median. They need the medium of these five values which would then be this fourth step. Now if you happen to give five then that would become the median because then you're in the middle there. And similarly if you happen to give three, that would become the median. If you give only a two or one, the median would still be three because that's the middle value. And so if the system does this for all of these 91 stations which is of course no difficulty for a computer, then it has basically a profile of the group as a whole. And then it can see what the most popular station would be at the moment. >> So, this maybe an approach has some interesting benefits that it seems like half the people are reasonably happy, but potential drawbacks. You mentioned this isn't exactly what MusicFX did. What did it actually do? >> Well on one hand they used a more complicated type of averaging median with sort of average square. But that is not the most interesting aspect, one thing they did in the beginning was to impose a threshold. Saying that no station will be considered if even one person hates it as in this case here. And they thought that was a cool idea because as you just indicated didn't want to make anybody miserable. >> Well, this is interesting because we had built a system called Polly lens that was basically the multi user version of Movie lens that's still out there. And we had to face the same question, and we decided to go against the idea of averaging or even median what we picked for Polly lens and I'm not claiming this was right. In fact we know it was wrong, was that we said that the score we would predict for a movie, was the prediction for the person who liked it the least. So that if there was a movie that five of us really would love then the sixth person would hate, and there was a movie that all six of us could barely tolerate. We'd go to the one that all six of us could barely tolerate. And it was interesting because what we learned In the process was that was a pretty lousy recommender. Because it meant that that person who liked things the least always drove the decision. And that the person who was willing to be grumpy and who just said I don't like much of anything, ended up picking the movie. And when we looked at how people make decisions in the real world, they often ignore the person who's always grumpy. Deciding that, that person is not going to be happy anyway, we might as well get with the rest of us like and left them be grumpy. This also raises the issue of manipulability. Whether people honestly disclose their preferences or whether they instead say something to try to get the effect that they want rather than be honest. How does that work with the system like MusicFX? >> Yeah, actually that came out very clearly as one of their first results. They had this system as I mentioned where if something is hated wouldn't be used. Sort of related to the misery approach you just mentioned. Now, what they found would be the people would explicitly choose that hate rating simply if they didn't especially like something. Not if they thought it would be some better station then they said they hated the current one, whether they hated it or not. And consequently this system soon became unuseful. You would soon get to a situation where no music at all could be played because someone would have said they hated it even if no-one really hated it. So that was a clear example of manipulation, in fact they changed that algorithm very quickly to avoid that problem. >> Yes I remember Joe McCarthy the person who led the effort of MusicFx telling the story about, the people in the gym who would exercise their power by rating something hated when it was playing so they could change the music. And watching when they've changed the system so that people who went to do that and tried it discovered it didn't work anymore. Which apparently led to some people being quite disappointed that their power had been diminished. We should point out that that kind of manipulation occurs even in a system like Doodle. We probably all know people who are very accommodating and scheduling a meeting and they will check. I could be available all of these different times. There are other people who say I really want to meet Wednesday between two and four. So those are the only times I'm going to check on the meeting form. And it requires a social process to figure out whether you respect those people's constraints. Whether you cut them out of meetings if they start being too difficult to schedule, or whether somebody has to apply social pressure to say, you know what? You need to provide more times, you're not playing as part of the team. So this whole problem of manipulation is a real one. I should- >> Yeah. >> Go ahead. >> Sorry, and you just give it a good characterization of what a mess it can be if people start manipulating, and even if people are aware of the problem and try to solve it by applying social pressure in the right, it's still quite difficult. You can't really force people to be honest, but the interesting thing is there are some aggregation algorithms which to force people to be honest. For example is median one has the advantage, that no matter how evil minded you are. You have nothing whatsoever to gain by specifying anything other than your true preference. Not to leave that mathematics as an exercise for your students but if you just look through the various different cases. You can see that you will never have anything to gain by specifying anything other than what you really believe. And that's the best thing because, and you don't have to worry about all those processing that you just mentioned. Of course it also implies that you have to convince people that it's true that they have nothing to gain by manipulating. The nice thing about the media, this is so simple people may actually pretty well see what is the case. >> So, while we talk primarily about the algorithms here, maniputability, calculation. We should point out that this is in a larger context that involves specifying preferences, decision making at the end. All of those aspects to come up with a reasonable use of a group recommendation. And some of these will come up later in the class when we're talking with pro approve from EPFL, whose done a lot of work in the listening preferences. But I know that you have some thoughts particularly about interfaces and ways of thinking more creatively about how people might express preferences. When we're dealing with groups of people, particularly groups of people who may already know one another. >> You and I talked a moment ago about the fact that, for example, when using Doodle, people will often save time by just not bothering with times that are obviously inappropriate. But there are more subtle ways in which people can take advantage of what they see other people to have specified. Why don't we have a look at a slide for this purpose? We saw this slide before, but here's a point we didn't mention. Let's suppose you are this new person here who's going to edit your preferences. And let's suppose, you don't really feel like having a meeting at lunch time on Monday, the 12th. But you sort of see that if you don't express willingness to take that time, there's not going to be any good time. So you figure, well, let me Just help the whole group out a bit to make it easier to find a solution so you actually specify that time as one when you're available. Okay this is a phenomenon of simulation that we noticed about ten years ago when I first started to study this sort of thing. And it turns out to be quite important, in fact just this year as the CSCW conference there was a study involving what the million Doodle users where a tendency was found that. If users can see what other people have already specified, the group's whole tendency will converge on a solution more quickly. Because people tend to assimilate their preferences in the way that I just indicated. >> And I think others have found similar things in areas like travel where. If you make the preferences of different people visible, then when one of them sees, my partner wants a beach and I want skiing. They can converge towards a compromise which might involve three days skiing and three days at the beach. Much more easily than if they're simply seeing the results of scores associated with each one of them. Well, this whole area of group recommendation is obviously an interesting and challenging one. And thank you for joining us to shed some light on the challenges and techniques and group recommendation. >> Sure, yeah glad to do it and feel free to visit my home page for more information. This was just a tip of the iceberg, and it's a good area for creative people to get involved in.