Recommender Problems for Web Applications
author: Bee-Chung Chen, LinkedIn Corporation
published: Oct. 1, 2010, recorded: July 2010, views: 10424
Report a problem or upload filesIf you have found a problem with this lecture or would like to send us extra material, articles, exercises, etc., please use our ticket system to describe your request and upload the data.
Enter your e-mail into the 'Cc' field, and we will keep you updated with your request's status.
In this half-day tutorial, we provide an in-depth introduction of data mining challenges that arise in the context of recommender problems for web applications. Since Netflix released a large movie ratings dataset, recommender problems have received considerable attention. The focus of this tutorial however is on web applications and we will cover topics that are significant extensions of those discussed in the context of Netflix contest. While the research pursued in the context of Netflix contest made great advances in “offline-research” (fitting model to historic data obtained from a recommender system), this tutorial goes beyond this and discusses important issues in “online-research” (the science of designing sequential schemes that serve items to users in an optimal fashion). At an abstract level, the goal of recommender systems is to display items from an inventory for each user visit to some website. Each display results in some response from the user (clicks, ratings and so on) that updates our belief about user preferences and results in better recommendations in the future. The data mining challenge is to construct a serving scheme or sequential design that learns user preferences through interactions with items in order to maximize some utility function over a long time horizon. For instance, a portal like Yahoo! maybe interested in constructing a serving scheme that displays articles to users visiting their front page to maximize click rates. The tutorial will begin with a formal definition of the problem through real life examples drawn from actual applications. We provide a detailed discussion of various scenarios under which recommendations may have to be made, including varying item pool size, dynamic versus static item pool, degrees of cold-start both for users and items, number of items to be selected for each display, user fatigue and so on. These scenarios go significantly beyond the classical movie recommendation problem. We then provide a comprehensive overview of state-of-the-art techniques in this area with detailed discussion of several open problems that we hope will stimulate further research. In particular, we describe time-series models, multi-armed bandit schemes, regression approaches, matrix factorization approaches, cold-start, similarity-based approaches exemplified through real world examples. We will end with detailed discussion of several technical challenges in this area.
Download slides: kdd2010_agarwal_chen_rpwa.pdf (4.7 MB)
Link this pageWould you like to put a link to this lecture on your homepage?
Go ahead! Copy the HTML snippet !
Write your own review or comment: