New objective functions for social collaborative filtering

Citation

Noel:2012:NOF:2187836.2187952,

 title = {New objective functions for social collaborative filtering},
 author = {Noel, Joseph and Sanner, Scott and Tran, Khoi-Nguyen and Christen, Peter and Xie},
 booktitle = {Proceedings of the 21st international conference on World Wide Web},
 series = {WWW '12},
 year = {2012},
 pages = {859--868},

Online version

pdf link to the paper

Summary

This paper points out three aspects which existing social collaborative filtering have overlooked; (1) learning similarities among users (2) learning restricted common interests among users and (3) learning direct user-to-user information diffusion. Then this paper proposes a unified framework that can consider the three aspects by introducing a new objective function.

To evaluate the accuracy of the proposed method, they created a Facebook application that recommends links to users every day and conducted live online user trials over five months, with over 100 Facebook App users and data from over 37,000 Facebook users. In the experiment, they compared the proposed method to several methods (including k-Nearest Neighbor, SVM, Matchbox CF, and several social CF) in live online recommendation tasks, showing that the proposed method has highest accuracy in "like" prediction task. In addition, they conducted user behaviors analyses by using real data they collected. They also got some feedback from users to improve social CF applications.

Problems authors point out regarding existing social collaborative filtering, and their solutions

• In existing social CF, it is fixed how to compute similarities among users, given some features for users. It would be more desirable if we can learn such similarities among various profile attributes, by using correlations of users.
  • To achieve this, they extended existing social regularization, which constrains the latent projection of users based on their social network information. They incorporated users' features in this regularization, by representing users by their features and mapping from these features to latent spaces.

• Existing social CF treats similarities among users only in global perspective. While two users are not globally similar, there may be sub-areas of interests which can be correlated to each other.
  • To deal with this co-preference, they added another regularization. They forced the following two components to be similar; (1) positive instances of preferences that are defined for 3-tuple (user, item, user) and (2) similarities among these three objects, based on their latent representations.

• Existing MF methods also do not model direct user-to-user information diffusion. If we observe a user always likes contents by a friend, it would be better if we can learn such a information diffusion.
  • To achieve this, they modeld direct joint features over user and items. Given a pair of (user (x), item (y)), newly introduced feature here is, for each friend (z) of the user (x), whether the friend (z) liked/disliked the item (y), and they used this feature in linear regressor. This allows the regressor to predict for any user (x) whether they are likely to follow their friend z’s preference for (y). They combined the output of the linear regression prediction with a MF based CF (matchbox prediction).

Strengths and weaknesses

Strength

The problems the authors pointed out regarding existing social CF are genuine to social media, but have not yet been fully considered. The authors propose a method that can solve the problem in a unified way based on MF. In addition, they actually created a Facebook application to collect user behavior data in Facebook. By doing so, they got rich features and conducted detailed analyses on users behaviors, too.

weakness

They manually set the weight for each objective function, and this might be time-consuming in practical situations. In addition, we are not clear which part of extensions actually contributed the increase of accuracy, because there are no systematic analyses. Thus, we cannot get much insight about users behaviors. We also cannot get much insight to improve the proposed method.

Possible impact

If they were able to publish data, it would have much more impact. (Actually, they cannot publish their data because of the requirement from the funding project.) Also, if they conducted analyses about how much each part of the extension contributed to the performance. If they did so, we could have insight about users' behaviors, or ways to improve existing social CF.

recommendation for whether or not to assign the paper as required/optional reading in later classes.

No. There are not much insight about phenomena in social media.

Related Papers

S. H. Yang, B. Long, A. Smola, N. Sadagopan, Z. Zheng, and H. Zha. Like like alike: Joint friendship and interest propagation in social networks.In WWW-11, 2011.

Study Plan

To understand some matrix calculation, I read some of the paper; K. B. Petersen and M. S. Pedersen. The matrix cookbook, 2008.