[SOUND] Hello, I'm Michael Ekstrand, and I'd like to welcome you to the Recommender Systems course and specialization. Our objective in this video is to provide you with some background to understand what a recommender system is, some of the history and background of these systems. We're not going to be getting into details in this video about the structure of the course or of the specialization. Those will be in the next video. Here, we want to set the stage for what it is that we're going to be talking with you about over the course of these courses. >> And I'm Joe Konstan. Welcome to Recommender Systems and let me take you back first a few hundred thousand years to the emergence of recommender systems as ants crawled around the face of the Earth. Well, this is isn't a history course, but we are going to spend a moment on the history of recommendation. But first we need a little bit of a definition as to what recommender systems are. They're systems that help people find things when the process of finding the information you need to make choices might be a little bit challenging, because there's too many choices, too many alternatives. We think of these in the modern day system by going to Amazon.com and having it recommend a book or going to Netflix and having it recommend a movie, but the ideas go much further back. If you've ever had ants in your home, you know that ants first spread out individually looking for food. But as soon as one of them finds food, it makes its way back to the nest, leaving behind the scent trail that the other ants soon follow. And then you have a stream of ants heading straight to your food and through your home. Those ants are exhibiting social navigation, a type of recommendation system where each of the ants goes out and explores a different part of the space, literally space, and when they find something that they think the community would like, they let everyone know about it. That's not limited to ants. Even before we had concepts like language and writing, cavemen almost certainly went through the same process. In fact, it wouldn't surprise me if there were some smart cave women, who, recognizing a new plant growing outside their cave, would say gee, I wonder if this one is good to eat, or if it's deadly and poisonous. Now, if you're starving, you have little choice. You grab it, you eat it, and you see what happens. But if you're wise and if you understand the lessons of recommender systems, you might wait a little while until somebody else hungrier than you, or maybe just dumber than you, goes and grabs that fruit and tries it. That person grabs the fruit, and if they start looking happy and it's delicious, this is wonderful, grab some for yourself. On the other hand, if they look sick or drop down, writhing in pain, you won a great, hard earned lesson, earned at the expense of someone who recommended to you not to eat that plant. Well, we can go much further than this. The concept of critics goes back thousands of years, people who would tell us what they thought we should see or not see, who are the great story tellers or poets, or artists, or actors. And of course, critics are there to help us make decisions without having to go through the long process of seeing everything that they saw. But let's jump forward a bit further into the more present age and look at concepts of information retrieval and information filtering. Information retrieval evolved in response to the need to be able to ask questions about a large collection of documents. Now much of the computing here was actually done because of large lawsuits that were being handled in the computer industry companies like IBM, but the same technology applies to libraries and their card catalogs or even to companies that are building indexes of the world wide web. The principles are the same. You have a static content base, or mostly static. We don't publish new books that often, compared to how often we read them. Or we don't publish new webpages as often as people navigate to them. But we have a dynamic information need. That information need is what we sometimes call a query and a femoral interest that we want an answer to. Because of this balance, we spend our time and invest it in indexing everything we can about that content base. We build up catalogs of the library. In the old days, we would have three different card catalogs, one by author, one by subject, one by title. If a book had multiple authors, we'd put multiple cards in that catalog, so that when you came in saying, I'm looking for a book by so-and-so, someone could quickly look it up, give you the answer, and then move on to the next query. There's lots of approaches to this, we're going to learn one of them that we use in our content filtering later in this course. It's one that's called TFIDF, which ranks the documents that we might be looking at, the books, the articles, the web pages, by how much they're about what we're looking for. But at the same time, ranks the terms by how good they are at discriminating among the different documents or alternatives. Now this information filtering is still used today, but as time moved on, we gained a new problem. We gained a problem of streams of information that were rapidly changing. Now you may get some of these streams, your email is coming in with, in my case, perhaps half a dozen or a dozen different messages coming in in an hour. You're trying to follow the news wire, and it seems that there's a new article posted every minute, but you don't care about all of the articles posted. You just want the ones that are most relevant to you. Information filtering saw that the assumptions of information retrieval were reversed. The information need here is pretty much static. I have a profile of what I'm looking for, I'm looking for articles on recommender systems. Maybe I'm looking for articles that mention my name. But the content base is dynamic, each article flies in quickly, and I want to know about it as quickly as possible. So in information filtering, I switch my effort. I invest my effort in modelling the user need. What is it that I'm interested in? Maybe by creating a profile by hand or learning it from my preferences. But sometimes, just looking at relevance is not enough. Collaborative filtering emerged as a reaction to the problem that sometimes there's so much content that you don't just want what's on topic, you want the ones that are really good on that topic. The first efforts of collaborative filtering were manual efforts based on the premise that key words were just insufficient. These systems based their filtering on a community of like minded users. System called Tapestry allowed people to mark up documents in a database saying >> Who do you think might like this document or what's good about it? Another system that was called Active Collaborative Filtering by Maltz and Ehrlich looked at the idea of helping people form distribution groups to get the good content out to the right people. Today, we have lots of distribution groups and some of them are use very effectively to get content to the right people and others you wish you can unsubscribed too. But perhaps you have relative you defend if you said, I just don't like your humor as postings anymore, but the idea is very simple. It's gee, if somebody could look at this for me, then maybe I don't have to screen through it myself first and that lead to the idea of Automated Collaborative Filtering. Automated Collaborative Filtering, which is the first system that became known as a recommender system started with the GroupLens Project. This work done in the early 1990s did Automated Collaborative Filtering for Usenet News. These are discussion lists filled with postings, which even back then and indeed still today. People felt were mostly not very good even if they were on topic, but buried in there where lots of gems that you wish you could get to without having to redraw all the stuff that wasn't good. The premise of GroupLens done by Paul Resnick, John Riddle and their students was that users who are reading news articles through GroupLens would rate the articles as they read them. They just put in a quick number for a one through five and users would be matched to each other to find other people who had similar tastes. When you came to the newsgroup to decide what articles to read, you would get a personalized prediction of which articles you would like or dislike and how much using a nearest neighbor approach where combined together the ratings of other people like you. This is what that looked like. This is a screenshot from one of the earlier GroupLens Usenet news readers. This is one called XRN and it's a screenshot from 1995. On the top right, there's a set of bars made out of characters that are the system's prediction of how much this particular reader would like each of the articles. This is showing a recipe newsgroup, rec.food.recipes and it's showing that this particular reader is told you'd probably be happier with Swedish flatbread than with minestrone. Why? Because other people who've agreed with you in the past have already looked at these recipes and they've told us what you think. By the way, when you're done reading this recipe, click a button on the bottom to tell us if it's a one through five or hit the keyboard one through five and we'll pass that information on to the next person. So, does it work? It sure does. In a trial of about 200 users using the group line systems for Usenet news, the ratings that people gave articles and the predictions that the system did or would have shown them were highly correlated. And more important, they were more highly correlated in the personalized case. The benefits of collaborative filtering were biggest in areas with the biggest noise like rec.humor, news group that's supposed to be filled with funny things, but too often wasn't. But there are still big benefits in groups like recipes where there was a particular benefit to personalization, because all the recipes were reasonable recipes, but they might not match your taste. So this was more accurate than predicting the average, more accurate than predicting the most common and particularly more accurate if you broke people down by news group and found their neighbors in the topic they were looking for. It also worked well in a way that meant something to users. People were significantly more likely to read the articles that the system said, they were going to like twice as likely and people like the system. Even after the test are over, they spend another year using the system just because they thought it was something that it was valuable and worth using, so that takes us to the commercial error of recommender systems. By the time we're in the mid and late 90s, companies were springing up left and right. GroupLens wasn't a loan months afterward, there was a system called Ringo and Homer from the MIT Media Lab that became Firefly Networks. The GroupLens system became the company Net Perceptions, Firefly became Agents Inc. Work was being done left and right, and folks went out, and got these things into commercial practice. This is a screenshot from Amazon with just one example of the literally dozens and then hundreds and then thousands of uses commercially, of recommender technology starting in the mid-90s and moving forward to today. You can't go to a retail site that's successful without them trying to guide you to the products that seem most relevant for you. But even if you're talking about catalog sales by telephone or any other place where there's a possibility of personalization and even in stores where we might group things together if not personalized for the customers in general, we're seeing the technology of recommender systems deployed pretty much everywhere. >> So, we like to spend a little bit of time walking through how some of these algorithms work. We're going to cover them in a lot more detail in subsequent videos in this in the next course. But to give you a taste, I'm going to describe how classic user-based collaborative filtering works using the example of recommending movies. So in this system, we have users and movies and the users assign ratings to various movies. These ratings are collected and stored by the system in its ratings database. The system then mines this database in order to find correlation between users and look for how much do different users agree on the movie that they like, and then sits, and waits for a user to come in, and request recommendations. When the user requests recommendations, the system uses its correlations to find a neighborhood of users who agree with this user on what movies are good. It then uses those neighbors's opinions to provide recommendations for this user. The idea being that if you agree on a lot of the things you've seen already, then the things seen by this person who agrees with you might be a good recommendation among the things that you haven't seen. So if it makes more sense to you to see this in a table rather than just a big blob of users, let's look at this set of ratings and say, I'm trying to decide whether I want to watch Blimp or Rocky XV tonight. So we're going to look for users that agree with me in our past ratings and we see that Ben, and I agree relatively well on a couple of movies that we've seen, and Nathan, and I also agree swapping out one of the movies. And we'll find out that Joe and I disagree quite strongly, but we also disagree fairly consistently. Maybe I might like something that he doesn't like. So then Pat and I agree, but Pat hasn't seen either Rocky XV or Blimp. So even though we agree, his opinions aren't that useful for figuring out what I should watch next. So now, we look at what do these users think about the movies. And we see that Nathan, who agrees with me likes Blimp, Joe doesn't, okay? Well, Joe disagrees with me a lot, maybe it's still worth watching. And for Rocky XV, Ben and I agree, and Ben really didn't like Rocky. So, it looks like tonight, I'll be watching Blimp. Now, these algorithms are not all there is to recommend our systems. They sit in a much bigger picture, so, we have some data store in the collaborative filtering case. It's a bunch of ratings or purchase behavior, of records of the way users have interacted with items. It can be the repository of content, it can be a database of movie reviews. There's many different kinds of data but we have some kind of data that we're using to produce the recommendations. The algorithms crunch these data and compute recommendations which are suitable for the current user and the current context etc. These recommendations are then presented to the users through some kind of the user experience. For example, on Amazon you've got recommendations of things you, related to the current item, things that might go with things in your shopping basket, etcetera. The way these are presented and then the way the feedback is collected from the user are all part of the user experience. And then we have the user themselves. So, the user receives these recommendations and acts on them in some way. This is determined by the way they think about, the way the process the recommendations. Psychology is working on studying this. Their input through the user experience is fed back into the database to help with future recommendations. And what we're going to talk a lot about algorithms in these courses. We're also going to be touching on some of these other issues, how the user experience works? Psychological aspects of how users perceive recommendations and what it means when they provide some of the feedback. So that you understand how these algorithms fit into the broader picture of what it takes to make a compelling recommendation experience. >> So when we think about recommenders? We could define them as tools to help identify worthwhile stuff for people. And we can break them down in one sense along the interface. Filtering interfaces take a stream of content like email or news articles and identify the ones that you want. I happen to use filtering interfaces all the time in my email because I get email messages from my department head. And if I spent all my time reading them, I might not have time left to realize that the former finance minister of some African Nation wants to help me transfer $6 million to my account. And I'd hate to miss that opportunity just because I was getting work done. So, I use email filtering interfaces, but we also have recommendation interfaces. These can be suggestion lists or top ten lists but more commonly they are placements. I go to a website like Amazon and a bunch of products are showing up there. They've been recommended. If I use my set top box for cable, Comcast shows me a few movies or shows that it thinks I might want to watch. Or if I go into a store, I may get a coupon printed out. That's a recommendation for a product that they think and perhaps hope I will purchase. Finally, we have prediction interfaces. This is what we saw in the use net example but there are a number of cases where what you want is some sort of a score like a predicted rating. If I'm looking at travel reviews, I might expect a site like TripAdvisor to tell me, this is a four and a half star, or in their case, four and a half circle hotel. It's a place we think you are really going to like. >> So as we go throughout these courses and we discuss more about recommender systems and their algorithms. It will help to have some vocabulary that we're going to use to talk about the systems. And a rating, we're going to use this term a lot and it's an expression of preference. These ratings come in two forms, there's an explicit rating which is something the user articulates. The user says, I like the Iron Giant four and a half stars. That is the user's explicit expression of how much they like that particular movie. But there's also implicit ratings that are inferred from user activity. If I'm watching Netflix and I start watching a movie and five minutes into it, I stop watching it, and I never come back. It's reasonable to infer from that, that I probably didn't like what I saw in that movie. There's many cases when we can infer from the user's behavior how much they likely like a particular product. A prediction is an estimate of this preference. So the user's preference, the rating is the preference information that we collect from the user. The prediction is the estimate of how much we think they're going to like items usually items that they haven't seen in the past. The recommendation is one or more selected items for the user. So the prediction says, we think that you're going to like to watch Frozen, four stars. A recommendation says we think Frozen would be a good movie for you to watch tonight, and maybe here are a couple of other suggestions. Content is attributes or text either in or about an item. So for a book or for a news article, the text in the article itself is content, so is its headline. For a movie, we can mine some things directly out of the content, but we also have a synopsis. We have who was in it. We have its title. We maybe have some user supplied tags, or the text of reviews about it. All of these things are content related to the item. And then when we talk about collaborative recommendation, what we're doing is rather than looking at the attributes, the content of the item itself, what we're looking at is the way that users interact with the item. Who likes this item? What else do people who buy this item buy. Those kinds of questions are collaborative questions because they look at what the community of users together do or don't do and the patterns in their behavior that emerge in that context. >> So along with this we have a variety of recommendation approaches. There are non personalized recommendations this is something that is just popular or everyone who's tries this likes it even if it's not popular yet. And stereotyped where we may say, gee this is popular if you happen to be a 15 to 18 year old female. Or if you happen to be a 35 to 44 year old Korean male. So both of those don't deal with direct personalization to the individual. But rather to the group or the population as a whole. We have product associations and this is the classic example of, gee people who liked this or bought this also like or buy that. You know if you enjoyed buying our Toyota Prius Wagon then we think you might also enjoy this solar charger, because it seems that other people have done so as well. These product associations are not personalized again to the individual. But they are connected to the context of a particular product that you are presumably looking at at that moment or that may be in your shopping basket. Content-based approaches start by learning what individuals like and building a profile. This is what we refer to as content filtering, or information filtering when it's about information. And then go out and figure out how the items that are available, or the products that are available, match up to that profile. And collaborative approaches learn what each person likes from the individual, and then use the experiences of everyone else in a community to make recommendations for other things that that person would like. >> There are several steps to designing and deploying a recommender. One is to collect the relevant data in order to inform a recommendation. A data on user's opinions, consumer's purchasing behavior, their experience with the item, a bit about the items themselves if we're doing content-based recommendation. You'd find the relevant data for particular purpose, if the user comes in and we define the data that's going to help recommend what they're going to like in the collaborative filtering example we saw that would be the other users' opinions that are similar to this user's previous opinions. Then we're going to compute the recommendations from this data, and finally, we have to present the data in a useful way that helps the user be aware of their options and evaluate the candidates that are being suggested. >> One final thought, tecommenders are a wonderful example of the techniques of Big Data. Recommenders are heavily about analysis and evaluation starting with this Big Data set of product opinions and perhaps product data, maybe even user demographics and other pieces of information we can pull together. We use that data to figure out the best recommendation approach. We then establish metrics that we're trying to optimize to and tune our algorithms to optimize performance against those metrics. And if we do everything right, those metrics are designed to maximize either some business goal, or user experience, or both. As we'll learn as we go forward in this topic, recommender systems are continually adopting new machine learning techniques to squeeze ever more bits of power out of the data that we have to provide better user experience. So as we draw this introduction to a close, we have to admit we don't know very much about you yet, but we're going to to guess that if you've made it this far, we highly recommend that you complete this course and this specialization on recommender systems. We look forward to seeing you in the future lectures.
