• by Laurie Sullivan , Staff Writer @lauriesullivan, May 10, 2012

A group of researchers at Microsoft and the CS Department of Technion-Israel Institute of Technology explored methods for modeling query and click behavior in Web searchers, as well as temporal characteristics in query and URLs. The research explores user behavior to predict changes in search queries, identify goals behind the queries, and analyze search results clicked on during Web search sessions.

The researchers' paper describes the opportunities found in studying human behavior on the Web through several models based on time series to represent and predict different aspects of search behavior. It examines how these models can learn from historical user-behavior data, and develop algorithms that address several properties seen in the dynamics of behavior on the Web, including trend, periodicity, noise, surprise, and seasonality.

Through predictive models, the researchers wanted to determine how user behavior changes over time. The group modeled behavior as a series of sessions in time, focusing on queries, URLs, and query-URL pairs as the behaviors. For example, they looked at the frequency of a query, and number of times a search result is clicked on for that query. The group presented different sequential activates and learning processes to show how to construct models that predict future behaviors from historical data.

Part of the research, a little geeky even for me, focused on learning to predict the behavior of users who search based on different temporal models through algorithms that learn from behavior by assigning a set of objects. These objects are given as examples for training the learning models.

Among the findings were  that time-aware modeling of user behavior can incorporate into many search-related applications. Query click prediction can be used to improve query suggestions to present the appropriate suggestions at the time of the query. URL click prediction can be used to improve re-crawling strategies, by focusing crawling efforts on URLs that are likely to be clicked. Query-URL prediction can produce better rankings more aware of intent.