Hello, I would say good morning or good afternoon but as you can see this is the part of the course on nonpersonalized recommendation. In this video, I'm going to be introducing this module talking about some of the roles of nonpersonalized and what we're going to call weakly personalized recommendation. And then in the subsequent videos, we'll be going into a lot more detail about different applications of nonpersonalized recommendation and how to compute these recommendations. So why do we want nonpersonalized recommendation? A lot of recommendations and a lot of the way we think about recommendation is personalized, such as Netflix suggesting what you might want to watch next. But there's a variety of situations in which non-personalized recommendation is still very useful. One is new users, where we know very little about the new user. You can't personalize, but you still want to be able to provide recommendations that they might find useful and actionable. Non-personalized recommendations can be very efficient to compute, things like how many people bought this item, is relatively fast to compute. And you still can provide a lot of benefit as far as sales, as far as the user finding the kinds of things they need so it can be a very good return on investment. There are a number of online communities that have built around access to common recommended news stories, such as Reddit or Slashdot. Where everyone sees the same set of stories that filtered at the top and then they comment about on them. And then there are applications in media where personalization is impossible. And non-personalized recommendation, as I said and as we'll see throughout the rest of this module, can be a remarkably effective tool. To provide a little bit of introduction, recommendation has long predated the kinds of digital recommender systems that we're familiar with today. Print magazines and newspapers have had book, and movie, and music reviews that served as a form of recommendation, or a form of prediction. Should I go see the new superhero movie this weekend? The New Yorker magazine has their Goings On About Town section that highlights a couple of shows that are in town, that are in New York in a particular week, that you can use as a form of recommendation for what to see if you're in New York. And the Michelin restaurant guides provide recommendations for good restaurants in various cities around the world. Each of these is editorially selected, so there is a set of editors that make the expert judgments on how highly to rate a particular restaurant, what shows to include in the page, etc. Another historical print recommendation is a book that was published in the early 1900s called The Negro Motorist Green-Book. And during the segregation era in the United States, when not all restaurants or hotels, or shops were available to all people, this book particularly listed accommodations that would be friendly to African-American travelers as they traveled across the United States. They could find somewhere to eat, somewhere to stay, somewhere to refuel their car. There are also print recommendations that are not based on expert editors so much, such as the Zagat Survey which you'll hear much more about in another video. And it aggregates opinions about restaurants from individual reviewers and surveys to produce a 30-point aggregate ratings on multiple dimensions. And it extracts text from individual reviewer reports to compile a final textual review as well, which produces something like this in its current web incarnation. Now, we talked in the previous module about the difference between implicit feedback and explicit feedback. The Zagat surveys aggregate explicit feedback from reviewers, looking at how much the reviewers says they like the food, the decor, the service, etc. Some non-personalized recommenders aggregate implicit feedback, such as how many times is a song or an album played on the radio, sold on iTunes, sold in record stores, etc. And the Billboard top 100 charts aggregate these kind of data in order to produce charts of the currently most popular songs and albums in the United States. There are many other examples of this kind of recommendation. The rating and review summaries on e-commerce sites such as Amazon. Box office charts list movies by their current sales. Most news sites show some kind of a most popular or popular now list, such as this one taken recently from the Atlantic. Moving beyond the non-personalized kinds of situations, sometimes we know a little bit about a user, such as their zip code or location or other demographic information, that we can use as a first pass to start to personalize. We can also do a little bit of weak personalization based on what item the user is currently looking at. If they're looking at a particular digital camera, they might be interested in related products, such as cases, screen covers and memory cards, that are compatible with the thing that they're currently looking at. So throughout the rest of this module, we're going to see how do you compute summary statistics to do non-personalized recommendation. We're going to look at how to do these weakly personalized recommendations based on demographics and stereotypes. How to identify related items and how to build these into compelling and useful user experiences. Throughout this, we'll also have a few different assignments for you. There will be an assignment to compute basic kinds of recommendations using a spreadsheet such as Excel or Google Sheets. There will be a quiz over the topics that we present in this module. For those of you taking the honors track, we also have a programming assignment, where you will implement some of the algorithms that we describe and see the results. And so you can actually see recommendations coming from a real data set using the non-personalized and weekly personalized recommenders that we're describing in this module. The learning objectives we have for this module, by the end of this part of the course, we want you to understand the value and the drawbacks of non-personalized recommendations, so you can know when might you want to use one. And then when do you need to start looking at using more sophisticated recommendation techniques? We want you to be able to compute personalized and weekly personalized recommendations using a variety of techniques, and to be able to design a user experience around some of these simple aggregation and recommendation algorithms. Thank you, and we'll see you in the class discussions.
