Coast to Coast Seminar Series Fall 2015: "Recommender Systems: Exploiting Big User Feedback Data for Personalization"

Abstract

With the dramatic growth of internet-based services, increasingly large numbers of items, such as web pages, videos, products etc. become available to users. To cope with the information overload, users demand a personalized selection of items that are relevant to their needs, but are typically not willing to spend much effort on specifying their personal needs. This has motivated the development of recommender systems, which automatically learn the user needs from available user feedback data. Feedback data can be explicit, e.g. ratings of movies, or implicit, e.g. clickstream data which records parts of the screen a user clicks on. The idea of Collaborative Filtering (CF) is to exploit big feedback data from many users to make recommendations to each individual user, assuming that users that have provided similar feedback on some items are likely to have similar feedback on further items. In this talk, we will briefly introduce matrix factorization as a state-of-the-art method to implement CF. We will then present two of our own research projects that explore the potential of social media for improving the quality of recommendations. First, we will show how to leverage a social network among users and the well-known effects of social influence and selection. Second, we will present an approach that also considers textual feedback data, i.e. product reviews, to create more fine-grained recommendations taking into account that user feedback may be based on different aspects of an item. While recommender systems have been successful in many personalization tasks, they come with their own risks. We will conclude the talk with a discussion of privacy issues, in particular in the context of social networks, and the so-called “filter bubble”, the phenomenon that due to personalization users do not get exposed to information that could challenge their worldview.