Welcome back. In this course, we're going to move away from the simple popularity and the content based approaches to look at a version of recommendation called collaborative filtering. In collaborative filtering, we ignore the user and item attributes. We don't care what's in the item. We don't care who the user is. We only look at the interactions between users and items. And we mine patterns from these, such as looking at what people like you also bought. This has two significant benefits. One, it allows us to develop algorithms that operate independently of the characteristics of the particular items being recommended. It also allows us to mine interesting pattern, such as other users who also bought all of this Red Sox paraphernalia, really like the clam chowder. >> So the key concepts in this course, first, nearest neighbor collaborative filtering. As a concept, as you've just heard, it's based on the idea of finding a set of people like you, and understanding what it is that that set of people like you also like. We'll start with the user-user collaborative filtering algorithm, an algorithm that directly forms neighborhoods of other people. And we'll explore topics such as the formation of neighborhoods and the tuning parameters that make this algorithm work when it works particularly well. We'll also look at alternatives for historic agreement as ways of building neighborhoods, things like building neighborhoods based on social relationship or trust. We'll then transition into item-item collaborative filtering. Where we pivot things around and look at relationships among items defined by the users who have experience with each of the items. We'll deal with issues such as unary data. What happens when we only know about likes, but not dislikes? We will look at hybrids and extensions that can bring new information into item-item collaborative filtering. And we'll be talking about the practical implications of item-item collaborative filtering, including how this algorithm really got its start in commercial use. Finally, we'll pull together a bunch of advance topics, including the challenges of getting started with a recommend your system when you don't have much data yet. Recommending for groups. Explaining recommendations, and threats and attacks when people might try to undermine a recommender system. >> This course is structured around these algorithms, with the first two weeks spent on user-user collaborative filtering. Where we'll have an assignment to do basic user-user collaborative filtering in a spreadsheet, an Honors Assignment for those of you taking that track, where you'll implement user-user CF in LensKit, and then a quiz over the concepts. Weeks three and four will be similarly structured, looking at item-item collaborative filtering and the advanced topics. Again, with an assignment, a programming assignment and a quiz to test your general knowledge of the subject. >> So, let's go on into the course.