Welcome back. As we wrap up our discussion of the Item-Item Collaborative Filtering algorithm, wanted to take a moment out to just bring together the real world experience of the strengths and weaknesses of this algorithm for recommending in different contexts. So let's start by going back to the two ways of thinking about item-item and the intuitions behind it. One way of thinking about Item-Item was as a more efficient implementation of User-User. At the time that these algorithms were being invented, we were thinking about the fact that we had this matrix that was painfully taller than it was wide. We had lots of users compared to the number of items. And yet, all of our comparisons were row to row, and that led to both less data in each comparison and more comparisons in order to do User-user collaborative filtering. And so, part of the intuition was if we could somehow flip the computation to work on the columns, the item, rather than the rows, the users. We might achieve the same goal, but achieve that goal more efficiently. The other intuition, perhaps the intuition that brought Item-Item to the folks at places like Amazon that were developing these in commercial recommenders early, was Item-Item as an aggregating Product-Association Recommender. Now we're already doing things like, gee, for this product, here's the most common other product. What if we just extend that notion to say well, what about for all of the products I've ever had together, what are the most popular products that go with that? And it turns out that computing that, doing a pure probabilistic approach might not be the easiest thing. But you could very quickly get into this, well if I have item correlations, I can take the items that best matched each of my products and then pull them together and emerge with an algorithm very much like our Item-Item Collaborative filtering algorithm. So one of these starts with the idea that I like User-User collaborative filtering. I like the idea of hearing what other people like, but I want to do it more efficiently. The other says, I love the idea of saying, well, if you're buying this you might want to buy that. But rather than limit it to what you're buying right now, what if I extend it to everything you've ever bought before? And as those two intuitions come together, we get the reality of the Item-Item algorithm. The early experiences, interestingly, were different depending on the domain in which this was deployed. All reports were that Item-Item was very successful in commercial application. Amazon made a big deal about it, had gone out and, did a bunch of work in that space. Other people were using Item-Item algorithms. In fact, in our work with the company and that perceptions, we moved to the Item-Item algorithms for performance reasons, specifically because the companyies wanted it. And they were generally delighted with the results, including results that were testing how often the things we recommended would actually end up being purchased. Around that time we deployed Item-Item in MovieLens when we switched from using our original User-User implementation, and it was a big disappointment. So what happened? Well, it's interesting because Amazon was happy the recommendations worked and MovieLens users complained. They said that the things we were recommending were too obvious, and that the recommendations and predictions weren't bold enough, and they were used to things that were bolder and less obvious. In fact, this is a large part of what led to the line of research in metrics to evaluate beyond accuracy that we'll be talking about in the next course in this specialization. So what happened from a very simple perspective is that Item-Item is not the same as User-User. It's very difficult if you go through the items that are commonly co-purchased with the items that you've already purchased to discover things that are very different from what you've already consumed. That's just not likely to be commonly co-purchased with these items or commonly purchased by the same user. User-User on the otherhand, elevates items that a close neighbor loves, even if there's very little evidence behind the fact that you would like that. So, if I have my top 20 neighbors and one of the top ones in that has gone off and I've mostly looked at comedies, but they see this sort of dark animated documentary. I don't know if you can do a dark animated documentary, but we'll go with it. And they loved it. That would be enough for User-User to recommend it to me because it has no evidence against it. Nobody else in my neighborhood may have ever heard of this movie so all the evidence is positive. User-User was built on the notion of social word of mouth and that if somebody says it's good, that's good enough. And Item-Item was built on the idea of let's aggregate the evidence and find the things that there's a lot of evidence supporting. Another consequence of this is that because Item-Item always was looking for a lot of evidence, its predictions tended to be less extreme. User-User had a lot of things that were given a five and frankly, a lot of things that were given a one. Because, those were things that only one or two people had experienced within your neighborhood. Item-Item, I never showed you those items, because it didn't have enough evidence to put them into your recommendations. With more data, you tend to get a little less extreme or perhaps more conservative. So what did we learn? Item-Item is faster and it's more stable for domains where there are more users than items. That stability is wonderful, it makes it possible to precompute and to store Item-Item correlation and do this system with a model that runs really quickly. But it's also substantially more conservative, which is good in many domains. If you're using up your opportunity to find something somebody's going to buy, Item-Item's probably a pretty good choice because it's something somebody's likely to buy. It's good in a consumption task. It may be frustrating, however, for people who think of a recommender as on its own, a source of entertainment and browsing and they're trying to discover things they would never have come up with before. And so knowing that, you can use how you apply your algorithms in an appropriate direction. So what next? One of the problems is easy to address. The fact that the predictions are more conservative is something that can be re-scaled. OrdRec is an approach for mapping a ranked list onto arbitrary prediction distribution. It's easy enough to say well gee, if you want 15%, 5 stars, we'll just make the top 15% into 5 stars. And matrix factorization techniques, which come in the fourth course of this specialization, present an interesting and different point in the design space where it may be possible to once again do a little bit more discovery. Though again those techniques do not reach the point of likely recommending, with very much accuracy, things that you have little evidence that you'll like. So with this, we're wrapping up our discussion of Item-Item Collaborative Filtering and how the algorithm works. As you go forward, you'll see a set of exercises including a quiz and the programming in spreadsheet problems that will help you pull together your knowledge of Item-Item collaborative filters and how to use it. We'll see you there.
