This paper uses data from Twitter on “follow” choices to explore the use of context for recommendation systems. Context here incorporates such things as usage patterns and rating behaviors, and factors that users may inherit from membership in social groups. The authors remark that even for a single user, the context may vary, for example, depending on the time of day.

Hurrell and Smeaton collected data from Twitter users in the Dublin, Ireland area. Of the data fields associated with the tweets, the authors selected 37 features, including the source location and language of the tweet, and descriptors determined by the tweeter’s name, such as the length of the screen name. For each user, the authors ran the features through a support vector machine (SVM) to learn how each user benefited from the feature. The results from the SVM show that no one group of features works for all users. Follower count has the best predictive value, selected by the SVM for 35 percent of users. On the other hand, the system does find context features for users that can be used to improve recommendations.

The paper serves as a good introduction to how large datasets obtained from the web can be used as research tools. For the specific case of recommendation systems, the results show that context can be a useful tool for personalization. The authors suggest that developers might consider creating a “context profile” that could accompany users in the cloud.