Welcome back. In the last few videos, we've seen how to implement item-item collaborative filtering, how to apply it to different kinds of data. In this video, we're going to talk about how to extend item-item collaborative filtering with additional concepts that can be useful in different kinds of domains. Item-item's a very good algorithm for extending because it's built of a number of simple parts with well defined interfaces. We can modify these individual parts, we can swap them out and recombine them in various ways. And it's fairly easy to understand what these kinds of extensions and recombinations and reconfigurations are doing to the algorithm. So I'm going to walk through a number of examples of ways that we can extend item-item collaborative filtering. You'll see a lot more of this kind of thing in the last course of the specialization, where we talk about hybrid algorithms, but there's a number of things that we can do right now with item-item. One example is incorporating a concept of the trustworthiness of users. If you remember from the interview with Paul Resnik in the first module, it can be useful to look at the trustworthiness, the reliability, of a user when we incorporate their ratings into the overall collaborative filtering structure. So, one way we can do this in item-item is if we have some kind of a global reputation score for a user. We can weight users by trust before computing item similarities. What can this look like is when we compute our weights, wij, We can add in a reputation score, I will call it rho, for each user, For the normalized rating. And then similarly, we can, in our denominator we will, Include our rhos. So that each users who have high trust, high reputation, have more impact in the recommendations. You can see this in Massa and Avesani's 2004 paper on Trust-Aware Collaborative Filtering. We can also extend it with notions of importance for items. Between two similarly similar items, we might want to consider one that has, is somehow more important to the overall space of items, to be a better recommendation. One application of this is when we're recommending research papers. It can be useful to incorporate notions of how highly cited a particular paper is, in addition to how similar it is to another paper. This is similar to the idea of scoring web pages by how central they are. And so the solution, one way to do this, is to weight the paper vectors by using PageRank of HITS or another scoring system that computes an authority or an importance score for an item. We could also take the item-item structure and replace the weights with something else entirely. For example, the similarity between item document vectors and content-based filtering, effectively building item-item content-based filtering instead of collaborative filtering. The core structure of the item-item algorithm doesn't care how its similarity is computed, and this can work relatively well. Finally, we can directly derive weights, instead of computing them from a similarity. By looking, using machine learning or optimization approaches to find coefficients that directly minimize, for example, the squared error of predicting ratings. We're going to see more of this in the next module applied to a different recommendation model. But there are a variety of optimization algorithms that can be used to directly learn these coefficients. To conclude, item-item collaborative filtering is a very flexible and versatile algorithm. You can build many interesting and different kinds of recommenders by reconfiguring it, adding additional concepts, reweighting vectors, and those kinds of things.
