Predicting User Interests from Contextual Information

            This paper present a systematic study of the effectiveness of five variant sources of contextual information for user interest modeling. The five contextual information sources used are: social, historic, task, collection, and user interaction. This study focus on website recommendations rather than search results. This research evaluate the utility of these five sources, and overlaps between them, based on how effectively they predict users’ future interests. The results demonstrate that the sources perform differently depending on the duration of the time window used for future prediction, and that context overlap outperforms any isolated source.

This research uses a systematic, log-based study of numerous contextual sources for modeling user interests during web interaction. The core task for any user modeling system is predicting future behavior, and evaluate the informativeness of different sources of contextual evidence based on their

informativeness for predicting users’ future interests at different temporal durations. Assume that the user has browsed to a web page and the task is to leverage context to predict their future interests. The use of the current page and five distinct sources of context are evaluated:

(i)        interaction: recent interaction behavior preceding the current page.

(ii)        collection: pages with hyperlinks to the current page.

(iii)      task: pages related to the current page by sharing the same search engine queries

(iv)      historic: the long term interests for the current user.

(v)       social: the combined interests of other users that also visit the current page.

This is the first study to systematically assess contextual variants for user interest modeling. The research also study the use of overlap between sources as a stronger source of contextual signal. After that the performance of contextual variants depends on the time duration used to represent future interests, and overlap between contexts yields more effective interest models than any model itself. Understanding which sources and source combinations best predict future user interests is critical for the development of effective website recommendation systems.

The primary source of data for this study was the anonymized logs of URLs visited by users who opted in to provide data through a widely-distributed browser toolbar. These log entries include a unique identifier for the user, a time-stamp for each page view, a unique browser window identifier, and the URL of the Web page visited. In order to remove variability caused by geographic and linguistic variation in search behavior, but only include entries generated in the English speaking United States locale.

This research studied the effectiveness of different sources of contextual evidence, and their overlap, for user interest modeling. The findings of our study suggest that the best-performing contextual sources are dependent on the duration between and the end of the prediction window. This has implications for the systems that use contextual information to support post-query navigation and general browsing behaviors. For example, these systems must not treat all context sources equally. Weights should be assigned to each source depending on whether the system is recommending web pages that are relevant to the immediate situation, the current work task, or the user’s general interests. The contexts as defined could be implemented using server-side lookups (task, collection and social) or client-side code (interaction and historic).