[MUSIC] Let's start by discussing some areas in which recommender systems are playing a really active role behind the scenes. And as we're gonna see in this discussion, depending on the specific application different aspects of the objective we're trying to optimize are gonna be important. But before we start talking about recommenders systems, it's really important to talk about the idea of personalization, because personalization is transforming our experience of the world. As we talked about in the clustering and similarity module, there are lots and lots of articles out there. In lots and lots of webpages. And we can't possibly be expected to browse all of them. So another example here is YouTube. So YouTube, it's quoted that there are roughly 100 hours of video recorded per minute. And the question is what do I care about. You know, I go on YouTube and I wanna watch some video that's of interest to me. So this is an example, clearly, of information overload. There's just way too much content there for me possibly to be able to go and for me to find what I'm interested in without somebody helping that content come to me. So browsing, the traditional form of browsing, is history. There needs to be some way in which the content that's relevant to me is automatically discovered. And so this the notion of personalization, where we wanna connect users like myself, who goes to YouTube and items which, in this case, are gonna be videos on YouTube need to somehow be connected. Okay, so, personalization is gonna play a key role in this notion of recommender systems and now let's talk about some examples where recommender systems are very important. So one very classic example that we mentioned a little bit earlier was this idea of Netflix where, here people go and they watch videos. And Netflix has this goal of trying to make suggestions of which movies or TV shows might be of interest the person who's come to this site. And the question is, how are we gonna go about doing these recommendations. Another example, again going back to something we talked about earlier, is Amazon making product recommendations. So you go on to Amazon, you purchase some product, and you'll see on the site that it makes a recommendation of other products you might be interested in. But one important thing to incorporate in these recommendations is the fact that not only do you have to take into account the interest that the person had in this one session. So for example here we're showing that somebody bought a book about websites and might be interested in another book about web applications or for example my most recent purchase on Amazon were shoes for my son but the interest that I have in making that purchase of shoes isn't my only interest. I'm not only looking for shoes for a one-year old. There are lots of other reasons why I go to Amazon and make purchases. And if you look at my history of purchases you're gonna be able to make much better recommendations for me than just based on a single session. And likewise, recommendations might change over time. So there's interest in making recommendations, for example, what I might be interested in purchasing today. So if you look at my purchase history a year ago, I was buying a lot of newborn products. So newborn baby products. But that's probably not something that I'm very likely to purchase today. So the recommendations that Amazon presents to me today have to adapt with time. So just as on demand video with personalized recommendation has really revolutionized how people watch movies and TV shows, likewise there are a lot of websites that provide streaming audio with personal recommendations. However, in this case unlike thinking about on demand video here we want one song to play after another. And what we want is we want some coherent stream of songs. So songs I've liked I wanna play similar songs. I don't wanna rapidly switch between for example some cafe Indi songs, all of a sudden playing a heavy metal song. But at the same time I don't want a song I just heard to play again and again and again. So I want some sense of recommendations that are coherent but I also want them to provide a diverse sequence of songs for me to listen to. Okay, well another critical area where recommender systems have played a very active role is in social networking. So for example on Facebook there are tons and tons of users. And we wanna form connections between these users. So for example here might be a graph of connections between users on Facebook and maybe I'm this pink node right here, and Facebook wants to recommend other people I might be interested in connecting with. In this application and it's important to note that both the users and the quote unquote items are of the same type. We're both people. So when I'm a user on Facebook the things that are being recommended to me, the items, they are other people. So we're gonna end up with users and items being exactly the same type in this application. But the recommendation system that we've talked about have really focused on online media. But more and more, people are realizing other area's in which recomender systems can play a really important role. And just as one example, we can think about what is call drug-target interactions. And here, maybe we have some drug that's been studied, for example, let's talk about aspirin. It's been well studied as a treatment for headaches. But what if it's discovered to have some other possible use? So for example, for blood thinning, for heart patients? If we can find these types of relationships, if we can repurpose this drug for some other treatment, then that could be really useful, because it's really quite costly and lengthy process to get FDA approval for a completely new drug, but if we can take a drug where the types of side effects and the possible risks associated are already well known and well studied. Then it's a lot easier to get approval for treatment with some other condition, so this is a case where we might say, if you like Aspirin for headaches, you might also like Aspirin for your heart condition. So recommender systems are playing an active role in medicine as just an example of the diversity of applications where we see these types of systems. >> [MUSIC]
