So now we'll take our prediction, a step further. And now we'll actually look at similarity. So before we said that we were just looking at, like user movie. similarities in a sense.where there weren't really any similarities. We're just looking at trends across the users and see how bias he was in his ratings or how good a movie was relative. Now we'll actually look at seeing how similar 2 movies are to one another, or equivalently really how similar 2 users are to one another, very similar technique to do that. so the main idea here in the the technical term for it is collaborative filtering, we really look for patterns across the table rather than just treating the rows and columns in isolation, we look for patterns across the table. And we'll look at one of these simpler methods. one of the easier ways to do collaborative filtering. There's many different ways to accomplish it. the first thing we'll look at is user to user. Right so let's just look at 2 users movie taste for instance. And suppose we have 2 users Anna and Ben, okay. So there is and there is 4 movies here, or 4 of them that they both rated Good Will Hunting, A Beautiful Mind, Lion King and Aladdin. So Good Will Hunting, Anna and Ben both gave thumbs up, Beautiful Mind they both gave thumbs up. Lion King they both didn't like. Aladdin they both didn't like. And Jurassic Park, Anna gave the thumbs up to. So, the question is can we make a prediction for what we think Ben would think of this movie? Well, him and Anna really gave the same signals for every other movie. Right? They either both liked the movie or they both didn't like the movie so, it seems like. Whatever Anna does is what Ben does as well, because they have very similar movie tastes. We would say, well, Ben would probably like this movie. Similarly if Anna didn't like this movie, didn't like Jurassic Park, we'd say that Ben probably didn't like it, right? So it works either way. And, so now this would also work for instance, and we'll look at this in a second with a movie example, but, suppose now instead keep Anna thumbs up, thumbs up, thumbs down, thumbs down. Suppose that Ben was thumbs down, thumbs down, thumbs up, thumbs up, right. So now they have opposite taste. Okay. So they're entirely dissimilar. Now that's still helpful, because if Anna likes this movie, than we expect that Ben will not like it because they have very opposite tastes. That's still helpful, and we'll see how we quantify helpfulness each direction in a minute. but that is still helpful then, we just do the opposites. And so again this either strong positive correlation which we just illustrated, or a strong negative correlation. Neither one is useful, okay we just don't want non no correlation so we don't want variations of whether they agree or disagree. And now the similar or dissimilar whichever ones are, whichever ones are talking about if Anna and Ben are very similar then we'll call them neighbors, and similarly if they are dissimilar we also call them neighbors. Aand that's what we use when we develop what's called the neighborhood Method. And we'll look at how we developed the neighborhood predictor throughout this, the rest of this lecture, really. And now additionally, we can also do movie to movie. [INAUDIBLE] Now let's look at a bunch of users and we'll see what they think of 2 different movies Good Will Hunting and A Beautiful Mind. So now, Anna likes Good Will Hunting, didn't like Beautiful Mind too much. Ben didn't like Good Will Hunting, liked Beautiful Mind. Charlie didn't like didn't like Good Will Hunting, liked Beautiful Mind. So now, we'd say these movies are very dissimilar. Alright. These users were similar, but the movies, now these movies seem to be dissimilar. So, Dana if we know that Dana liked Good Will Hunting, we say it's probably she did not like Beautiful Mind. So, it's just the, or we are just cutting the table differently. Here, we are looking across the columns because we aren't to get data for how 2 different users responded differently to one another. So, user to user similarity because we are comparing 2 different users. And here we're looking down the rows because we want to look at 2 different movies and get the most data possible. See how users respond to the different movies and, and so we can predict Dana's rating for this movie, hopefully. And, similarly if Dana didn't like this movie, we wouldn't think that she would like Beautiful Mind. So it works either way. We'll focus on the movie to movie similarity, actually here.
