Welcome back. In today's interview, I'm pleased to bring you Jen Golbeck. The Director of the HCI Lab at the University of Maryland and also an Associate Professor at that university. Who's an expert in trust space recommendations and in social computing in general. In fact, before we go on, I'm going to ask you to tell us a little bit about her most recent book but first we're just going to welcome her to the class and start right in on trust based recommendations. Welcome, good to have you here, Jen. >> I'm great, really happy to be here. >> So far our class has learned about recommenders looking at non-personalized recommendations like averages and product associations, a content based filtering, building up personal profiles. And we've just gone through user-user collaborative filtering. Can you tell us what trust based recommendation is and how it fits into or extends that framework? >> Sure, so trust based recommendations are really a variation on collaborative filtering where you're doing user-user computations. So, the most basic idea of a user-user collaborative filtering algorithm is that you compute some sort of similarity between users. And then that becomes a weight that you can use to give priority to users who are more similar to you. And to enhance how much you would consider their ratings of items. For trust-based recommender systems, you are doing a similar thing. But instead of computing similarity with other users, you're actually leveraging an underlying social network and trust ratings that users have given to each other, either explicitly or that you're pulling implicitly from existing data. And using that in place of a similarity measure to make recommendations. So, the idea is that instead of giving more weight to people who are similar to you, you're giving more weight to people that you trust. And that captures some more nuanced things than similarity gets on its own. >> So, how would you compute trust in one of these systems? Are there different models that work in different situations? >> Yeah, there's a few different ways to do it. When I started doing this work as part of my dissertation in like 2003/2004, we were working on models where you had a social network like you'd have on Facebook and users would actually explicitly rate how much they trusted other people in the network. And we were doing that as part of research and so people were willing to give us that data. But the fact is that most people don't go into Facebook or Twitter and explicitly say how much they trust other people there. So, there are a handful of algorithms that will work to compute trust based on explicit ratings that people have given saying, how much they trust one another. And we'll look at a slide in a minute that goes through one of those algorithms because it's similar to most of them that are out there. But there's a number of other ways that you can do it. Sometimes, you'll infer trust by looking at something like similarity but that's actually much more nuanced. And you're computing a lot of other attributes between two people and their ratings. So, you get an estimate of trust based on similar ratings. But you don't require people to actually go in and say how much they trust one another. And there's some systems that are hybrid. And on top of that, there's two other ways of looking at this. One is in terms of what we would call reputation, and the other in terms of very specific personal trust. And we differentiate those where reputation is something that's global. So, if I want to know how trustworthy you are, I'm going to get the same answer as anybody else who asks. You have a general reputation. And a trust system that uses that would give the same the value to anyone who asks how trustworthy you are. The more personalized trust systems are actually going to give a different answer depending on who asks. And that can be really useful in situations, like, say politics. If I'm a very liberal person and you're a very conservative person, if I want to know how much to trust you about political issues, I'm probably going to get an answer that says not very trustworthy based on my personal perspective. But if someone who shares all of your views asks, the system may say that you are trustworthy because you have similarities with them. And so those systems really rely more on the social network and could have some additional interesting computational things. The slide that I've prepared for you and that you have shows how one of those local trust algorithms works that infers a different value for each person. So, we could go through that if you want. >> Sure, let's bring up that slide. >> Okay, so what we have here are two nodes. One is called the source and the other's called the sink. And that's basic network lingo. The source wants to know how much to trust the sink. So, in this algorithm, which is kind of a common base of a lot of algorithms that work this way. The source will ask its friends how much to trust the sink. So, if you click to the next, you'll see the source has a couple of friends. The one on the bottom might know the sink directly. So, if we click again, you'll see that there's an edge connecting that bottom node to the sink. And that node will return its value to the source. So, if you click again, you'll see we get an edge coming back. Click again and you'll see, we have a value that's stored down at that lower node. And that's going to get reported back to the source. So, the source has some information from its friend on the bottom. Now the friend on the top might not know the sink directly and so the friend on the top's going to ask its friends. So, if you click again, you'll see those friends show up. Some of those friends might know the sink, so if you click, you'll see an edge from some of them. Other of those friends might not know the sink. So, if you click again, you'll see that they just repeat the process. They ask their friends and this goes on until eventually you get some friends that know the sink directly. They'll grab their trust values and they'll pass it back to their friends. So, you can see some edges here and they start going back. Now when a node gets information from a couple of its friends, what it wants to do is take the values from those friends about how much they trust the sink. And then compute a value for how much it would trust the sink. And it does that with a weighted average. So, the friends it trusts more, their opinion gets more weight. So, you'll see a little formula, that's just a weighted average. And then this node now has a value for how trustworthy it thinks the sink is. And every node repeats that and if you keep clicking, you'll see that these values eventually get passed back, weighted in an average until they get back to the source which then does its own weighted average of its two neighbors. If you click again, that should show up underneath. And it ends up with a final score of how trustworthy it thinks the sink is. So, that's a simple version of the process. Now, if you click one more time, you'll see that there's actually a bunch of other nodes and edges included in this network that don't get factored into the computation. Because you use a method to try to pull out the most relevant nodes and the shortest paths between the source and the sink to finally get that value. >> All right, so while we have this slide up, a piece of this is clear. We're getting the trust value directly, when we have one, from our friends, when we don't. And if I trust my friend, I'm more likely to trust what they think of the person I'm trying to decide whether to trust. Does something go in later? Does the source update its trust value for those close friends? If it turns out they were right or wrong about their judgment of the sink, down the road? >> Yeah, that's a really interesting question. It's difficult. So, certainly, trust can change in any of these systems and the algorithms will update the inferred values. The problem is when you're doing an inferred value, you generally don't know which of your friends is telling you this because it's an algorithm that runs in the system in the background. And that's actually important because privacy is really important when you're talking about trust. And this was an issue when I started running these systems. I had some friends and colleagues putting in values. And I had one fellow graduate student and his advisor had put in a really low trust value for him. And he was so upset about this and you can think of reasons for that happening. But the fact is that most people want their trust ratings kept really private. So, generally, the system is not going to reveal what information came from what person. So, those trust ratings can remain private. So, you have to do a little bit of kind of personal investigation to figure out which of your friends is giving you good information and which isn't. >> Are there particular benefits and drawbacks that might work in certain applications here or other places where trust based recommendation just fails miserably? Yeah, there's places where it's good and there's places where it's really bad. But the thing that it's good at, is when the user, who's giving the recommendation, is very different from the average person. So, the common example I give from this, is looking at a movie recommender system. A Clockwork Orange is a great example. I hated A Clockwork Orange. Like it's a couple hours of my life that I wish I had back. But I understand why people give that movie high ratings. I don't ever want to see anything like that though. So, if you go into the system, if you have, say, a four star scale. Average ratings of that movie are like three and half stars because it's very important film, people give it high ratings. My rating is whatever the lowest option is in the system because I don't want to see anything like that again. In cases like that, collaborative filtering algorithms can have some problems. Because there may be people who are similar to me in a lot of ways. And in fact, people who like that movie tend to be kind of movie buffs like me. So, I'll get recommendations that are very high but I'm so different than everyone else that the collaborative filtering algorithm kind of fails miserably. But trust algorithms tend to do really well in this space. Because trust is certainly related to similarity, but it also captures things like how much do we agree on items that I'm really passionate about, that are important to me. And we've done a lot of studies that show that if you and I have one big disagreement on something I care a lot about, my trust for you drops way down. And so that means these trust based recommender systems are able to capture similarity on those important things better than a collaborative filtering algorithm can. So, they do much better in those cases. Overall, they perform about the same. But one of the big drawbacks to them is that getting accurate trust data is very hard. There are some systems where people provide it. But overall, you don't have a lot of it and so you're estimating it, you're pulling it from a network, and you're just, you have so much more bootstrapping to do than you would in a collaborative filtering type system. That it's hard to get enough initial data in the system to get really good results. It's a much more serious problem for trust recommenders than other recommender systems. >> I also think, it's sometimes just hard to explain. That the first notion when I hear trust based recommendation, I think, gee, I sort of trust my relatives but I wouldn't be willing to take either movie advice or restaurant or frankly even investment advice from most of them. I know them too well to trust them in certain specific domains, even if I'd be happy to trust them in others. And clarifying the whole idea that you build this trust network specifically for a domain seems to be one of those important parts that you have to get people over the hurdle to understand where the trust really comes from. >> Yeah, and so, that's a great point. I mean, the context of trust is so important. And people do have a hard time assigning trust ratings. And even if you explain them, we want you to think about movies, and how much would you trust a recommendation that comes from those people, they feel bad giving low trust ratings to people that they really like. There's just a psychological issue there, that you want to say, I trust this person about movies, even though they don't really reflect your views. And so that can put a lot of noise into the data as well. >> Well, and part of that is also the difference between trust and agreement. It might be that if I know you, even if we disagree on movies, if I know you know my taste, I would trust that if I ask you for a movie recommendation, it would be a good one because you know my tastes. But I don't trust your ratings of movies because we disagree. And the language is not quite subtle enough to get those differences out. >> Yeah, I think that's right, and that becomes a real problem in an automated system where you're not actually asking the person for a recommendation because they could give a good one. But you're just relying on their ratings. >> Great, so, are trust based systems out there in widespread use? Are there examples that we can point people to? >> Yeah, not as much as I would have liked when I was writing my dissertation on this. But the one real example of this is Epinions. So, it's epinions.com and it's a general review website, kind of like Yelp. But more for products than services in restaurants. They have a system called, Web of Trust, which uses trust ratings that you can assign to other people. And they propagate those through the network like I described. That company is now owned by eBay, actually. And sold for, I think, $650 million last time it changed hands. Lots of users, lots of activities. So, that's the one big example of where you can see these sorts of systems in use in real practice on the web. >> Wonderful, well before we go, tell us a little bit about your book. because while these students are here for recommender systems, I know many of them are more generally interested in social computing and analysis in social networks. And it's a nice opportunity to let them know about the work you've done. >> Great, yeah so, the book's called Analyzing the Social Web. And it's a textbook that I developed out of a class on social network and social media analysis that I taught at the University of Maryland. And I was teaching that both on the graduate level and as a core class for the undergraduates. And it pulls together things like graph theory and structural analysis, which you might do in a math or computer science course, along with sociological concepts like trust and tie strength. And really looking at how people interact and how you can study that and understand both networks and who's important in those networks by pulling in both computing, mathematical, physics concepts and then social science concepts. So, about half the book teaches this vocabulary of interdisciplinary ways of looking at networks and the other half is looking at case studies in different domains of how you can actually apply that and understand the networks better. So, I'm really happy with it because it brings a lot of different fields together. And it's a real interdisciplinary look at the topic. And it's written for people who have no background in the field. So, it's a good way to get started learning about how to do social network analysis. >> Well, wonderful. Well, thank you for joining us. It's been a pleasure to have a chance to talk with you. >> Thanks, this is so fun.
