Welcome back, in this video, we're going to talk about how to adapt the item-item collaborative filtering algorithm to work with implicit feedback or unary data. We've talked about how to do it over rating data, but it also works quite well over various kinds of unary data implicit feedback, such as link clicks, song plays, product purchases. But there are a few tweaks that we need to make for the algorithm. If you want more details on this, I refer you to the Carapace paper that's going to be linked to in the notes as accompanying this module. So first, we need to represent our data. We have the standard user item rating matrix but we need to represent this kind of purchase data or play data. What we can do, what's often done is using a logical 1/0 representation. So an item is 1 if they, rating is 1 is they purchase and 0 if they haven't. We can also use the purchase counter play count, so if we have data about users playing songs, we can represent, let's say the rating is the number of times the users played that song. And 0, if it's 0 if the user has never played that song. We don't pay too much attention to the zeros. We represent things as zeros that, there's some issues with that. For example, the fact that they have not played a song doesn't mean they don't like the song. It might just mean they don't know about the song. But, for the purposes of computing our recommendations with item-item, we don't worry about that. Also, in a large item space, most items are not going to be relevant to any given user. So, the zero's are probably irrelevant items even though some of them may be items to use you would very much like. There is also mean-centering we have a bunch of ones, mean-centering is not useful for thing to do. But instead, what we can do is we can normalize vectors to unit vectors. If we normalize the user vectors to unit vectors, then we get this notion that users who like many items, provide less information about the value of any particular item. We can also, if we have count data we can apply a log function to it, to decrease the impact of very large counts. Cosine similarity still works for computing the similarities between users or between items, works quite well. We can also use conditional probability as was done in the Deshpande and Karypis paper. So our cosine similarity, the w i j = cos( r ), usually normalized. Or w i j = probability that the user purchased i given that they purchased j. And there's some additional modifications that you can make to improve that. It becomes very, very similar to association rules. You can think of it effectively as plugging the association rules into the item-item collaborative filtering algorithm. It's not symmetric, so if you're doing the model build with conditional probability, you have to be very careful to make sure that you're using the probability function the right way. Because the probability of ri given rj is not equal to the probability of j given i. We also have to change how we aggregate the scores. For counts, etc., weighted average still works fine, but for binary 0/1 data, the weighted average doesn't work. And the cy, the cy, suppose that we computed the weighted average, But the ratings Are a bunch of 0s and 1s. So for the items the user's bought, you're summing up the weights multiplied by a bunch of 1s and then you're dividing by the weights again which equals 1. Which is not a very useful way to distinguish between different relevancies of items, if all the items have relevance of 1 or not a number because there are no related items. So instead what we can do is we say this S (u, i) = the sum of the j and the items purchased by you, w i j. We can also truncate this So we only consider the top k items. The neighborhood selection strategy is unchanged, we still pick the most similar items. So in conclusion, item-item basically works for unary data. There's a few things you have to change. You'll need to test the variance with your data, your context, and again, the material in the next course in evaluation is going to help you to do that.
