Electrical Engineering
      and Computer Sciences

Electrical Engineering and Computer Sciences

COLLEGE OF ENGINEERING

UC Berkeley

   

2009 Research Summary

Portfolio Effects in Collaborative Filtering

View Current Project Information

Tavi Nathanson, Ephrat Bitton and Ken Goldberg

National Science Foundation

We are considering the problem of recommending a weighted item portfolio, rather than an ordered list of items, in a collaborative filtering system. This problem is motivated by Donation Dashboard, launched in April 2008, which recommends a portfolio of non-profit donation amounts to users. The site currently uses the Eigentaste 2.0 constant-time collaborative filtering algorithm to collect ratings, cluster users, and predict ratings. These predictions are processed to generate a portfolio of recommendations that includes the user's previously rated items.

As of mid-September 2008, Donation Dashboard 1.0 has a database of 70 non-profit organizations, and over 2,000 users have registered on the site. We have collected over 33,000 ratings from all users. Data we collected has indicated that rating individual items on an absolute scale has significant dependencies on the order in which the items are presented. To mitigate this bias, we have been developing a graphical model and algorithm that use relative ratings to generate portfolios. This project is supported in part by craigslist.org and the Berkeley Center for New Media.

Figure 1
Figure 1: A user of Donation Dashboard 1.0 is rating a non-profit organization.

Figure 2
Figure 2: A user of Donation Dashboard 1.0 is recommending a portfolio.

Figure 3
Figure 3: Normalized histogram of sample relative ratings provided by the user and its corresponding graphical representation

[1]
K. Goldberg, T. Roeder, D. Gupta, and C. Perkins, "Eigentaste: A Constant Time Collaborative Filtering Algorithm," Information Retrieval, Vol. 4, No. 2, 2001, pp. 133-151.
[2]
T. Nathanson, E. Bitton, and K. Goldberg, "Eigentaste 5.0: Constant-Time Adaptability in a Recommender System Using Item Clustering," Proceedings of the 2007 ACM International Conference on Recommender Systems, 2007, pp. 149-152.