Welcome back. In this interview, we have the privilege of talking with Dan Cosley, Dan is a faculty member at Cornell University, who earned his PhD in our group, doing interesting work on recommender systems, he's now working more broadly in this whole social space of social awareness and social explanation, and all sorts of cool things. But, for today's interview, we're going to focus on a piece of work he did about a decade ago that got a lot of publicity. Because it looked at the whole question of what happens to those recommendations that people see, and do they believe them? And does it affect their opinion, having seen a prediction of what the system thinks they will think? So Dan, welcome, I'm delighted that you can join us. >> Thanks sir, how's it going? >> It's going well, so, I guess we should start with the question that motivated that research. Why don't you tell us, why did you go after looking at this question that ended up under the title, is seeing believing. >> Okay, so, first of all this was joint work with another student, Tony Lamb, and another research scientist, Albert, at the lab. And what we were interested in was people had. And people who had used MovieLens have been talking about when they were rating movies, they didn't want to see predictions, right? because they were sure that they were being influenced by those predictions, and that seemed like a pretty interesting question, and there were kind of good reasons to think about whether, think that an interface might actually influence you, right? I mean, there's been a lot of work where things, if you show on a wine label, you show it's a California versus a North Dakota wine. It'll all affect your perception of it, even if it's It's the same wine, right? Or if you give somebody just a random number in their head, then when they start thinking about like prices, or quantities of things later, it tends to anchor them on that number. I'm sure there's a lot of theoretical basis for why you might think about the fact that an interface showing you the prediction of what it thinks you like, might affect the way that you actually rate the item. And in particular, the one we chose, the way we chose to think about was based on this classic set of experiments by a guy named Ashe in the 50s. Where he asked people to come into a room, and they were told that they were doing a judgement task. So, say which of three lines on a screen was the longest one. And unbeknownst to the person in the experiment, everyone else in the room was a teammate of the experimenter. And so, after a while, they would start lying, and they would say that one of the shorter lines was actually the longest one. And often, the person who was in the experiment started agreeing with the group, even though he knew perfectly well, or she knew perfectly well that the line was shorter. And so, we wondered if these effects might happen in MovieLens, and other recommender systems where they show the rating, right? The predicted rating, right? Next to a place where you could actually rate the item yourself. And it's super common, because recommender systems need the data. >> So, what did you find? >> Well, so, we found that showing a prediction really does anchor people's behavior. So what we did, right, this was. Tony had this really great insight that a lot of movies people had already rated in movie one's, but a fairly long time ago. And so, what we did was we showed them movies that they had rated in the past. Right? And then, beside the place where, and we asked them to rate the movies, and beside the rating we showed them a prediction. All right, for some people, we didn't show them anything. We just showed question marks, like, movie one's computer prediction, yeah. Sometimes we showed them their actual prior rating of the movie, and then sometimes we took their prior rating and we added are lowered in a star from what they had originally rated. And so, we mixed these together randomly, so people didn't really know what was going on. And in fact, if you show people a prediction it anchored, if you show them the original prediction, they rate after their original prediction more often than you show them nothing. So, if you showed them nothing, they'll rate the same about 60% of the time lower about 20% and higher about 20%. If you show them exactly the rating that they had given as their prediction, it goes to 70, 15 and 15. And if you manipulate it down a star, it goes to 30, 60 and 10. If you show it up a star, it goes to 10, 60 and 30. So basically, whatever we show them, people tend to rate in that direction. >> So, that's a pretty interesting result, I mean, we know then that people are biased, a little bit. The way you modeled this, you could interpret it two different ways. You could interpret it as what happens if your recommender just isn't very good? That bad predictions might lead to bad data coming back in. But you also looked at it as the question of what if the recommender has been deliberately manipulated? And I know people have wondered this for a while. If I came in and, let's just assume for the moment that my goal was to get everyone to think that yours was the best interview in the class. I could've started the introduction by saying now we're going to get to what I think you're going to find to be the best interview, and just used my authority to do it. But I could've also- >> I've thought about wearing a five star t-shirt to try to influence people exactly in that kind of way. >> Exactly, little did you know we're on a nine star scale. >> Damn it. >> But we could also go the direction that what would have happened if I just planted four or five fake student accounts that each put in a little message saying, hey, this was really the greatest lecture. And so, why don't you talk us through what the logic was in the implications of what happens if somebody comes in and shills, or deliberately lies to try to make something look better or worse in a recommenders system. >> Right, so first of all, this happens all time online, from obvious spam where people are just posting links and comments. To more subtle stuff, they're posting fake reviews in Amazon and Yelp. But I've actually got some colleagues at Carnel including and Jeff Hinndencock and Claire who've been studying. Can you tell whether a review is fake or not, based on language that people are using? And it turns out that you can. And in fact, we'll get back to this. In MovieLens, people could tell that we were playing with movie ratings too. But the problem is that if you have this bad data in the system, you get downstream effect, sorry, it's a recommendor system that's using this data to calculate the recommendation. So, if I can get bogus high ratings on the system, then it's going to tend to be shown to more people. And it's going to be shown with a higher rating to those people. Which in turn, will influence some of those people. And so, you get this kind of vicious cycle of bogus in your system. So, is there any good news that came out of this day? >> No, [LAUGH] a little bit, so, we actually wanted to see whether people notice, right? And so, we, what we did was we asked some people to do the task, but didn't show them anything at all. We asked some people to do this task, and we were rating movies, but what we showed them, the manipulated predictions and the way we saw. And then the two groups when we saw some at the end of the experiment how well they liked MovieLens, right? Were the recommendations during the experiment useful? And was MovieLens useful overall? People who were in the conditional way of playing with the raise gave a significantly lower range, they liked their recommendations less, they liked movie lens last. And so, I'd say the manipulation was pretty course than what we did, right? We basically manipulated half of the movies for the experimental condition. But it was just 40 movies, and it was only 10 minutes. And somehow, we were able to significantly make MovieLens look worse. Now, that's not good news. But the good news is that that if people are reasonably sensitive to these things, and that it's likely that at least the companies themselves. Right, the recommender system company could also play around with the ratings, recommending products that they wanted to get off of their shelves, or recommending products with a higher profit margin. And to some extent, people can detect these things, and they'll react by liking you and going and doing something else. >> So, maybe the third or fourth time they start trying to sell you the extended warranty, you stop buying the television set there. Or maybe the other lesson for our class is if you're going to lie to people, save the lie for when it's really worth it. Don't waste. >> [LAUGH] >> You lies on small things. >> Yes, can I tell you something else about this too though? >> Absolutely. >> Just recommend our systems, right? This is, I mean, if you think about Google, right, and search results, they have an algorithm that's trying to pick the best things for us. And then the interface shows us the top ten, which then drives attention to those sites, which increases linking. So those, right, are, in Amazon, where it says people who viewed X, also view Y. Part of it is because in the Amazon interface, it shows you a thing where people who viewed X viewed Y, and so, it increases the probability that it's kind of a self-reinforcing effect. Or who's popular on Facebook, right? Good newsfeed algorithm, and how it decides to show stores to people who might actually affect the ways that you think about other people, and people's perceptions of Facebook as a system. And these are all kind of the same aesthetics. >> Yeah, and in fact, this is an argument that Google made from the beginning about why they didn't manipulate their algorithm. They said, look, people use us because we deliver the right things at the top of the list. If we start messing with it, people will notice and they won't think we're a very good search engine, and I think what your study is showing is that has the same effect in the recommender system. And we don't know if we people can distinguish between manipulated and bad. But in either case, they don't like getting bad results. Yeah, but I mean, also in either case, the interface is shaping beliefs and behaviors to some extent, right? And that's something, no matter what algorithm you put in, even if it's the best one or the most accurate one, it's going to tend to tweak people's opinions in that direction and have real consequences for the world, right? The rich get richer because- >> The rich get richer and/or the biased get more biased, right? So, MovieLens has this big dataset. And baked into the MovieLens dataset is the MovieLens algorithm and the MovieLens interface. And so, every researcher who ever has studied recommender system's has a little bit of influence based on the fact it came from these people through this process. And that's something that it's easy to forget about. We're doing kind of systems work or data mining work, or big data kind of work. >> As long as we're there, the early users of a system like MovieLens. Shaped the experience for the 100,000 people that followed. And this might be a message that, if you really want influence in the world, go find a system that's going to be popular, and be one of the early contributors of data, so that you can shape the direction the system goes. >> For sure. Think about Wikipedia too, right? A lot of the Wikipedia ethos and the way it is today came from that first 50 or 100 people who came along, right? Sort of shaped what it means to be Wikipedia. Whether a Wikipedia article is, what's allowed into it, not allowed into Wikipedia. The question, right, there's 50 million things to bet on. >> Right. >> You only have so much time. >> Always a challenge. Well, it's been delightful talking with you, and thank you for sharing these insights on whether seeing is believing, and what happens when manipulations work their way through the chain of perception, opinion, expression, and onward. >> Yep, thanks for having me, appreciate it.