Personalized Web Exploration with Task Models

Personalized Web search emerged as one of the hottest topics for both the Web industry and academic researchers . Unlike  traditional “one-size-fits-all” search engines, personalized search systems attempt to take into account interests, goals, and preferences of individual users in order to improve the relevance of search results and the overall retrieval experience. In the context of a tight competition between search engines and technologies, personalization is frequently considered as one of the technologies that can deliver a competitive advantage.

As long as the Web is getting closer and closer to becoming a primary source for all kinds of information, more and more users of  Web search engines run exploratory searches to solve their
everyday needs. In addition, a growing proportion of users, such as information analysts, are engaged in Web based exploratory search professionally by the nature of their jobs. It makes exploratory Web
search both attractive and an important focus for research on search personalization. Yet, only a very small fraction of projects devoted to search personalization seek to support exploratory search.

The most typical usage of the profile is to rank information items. Filtering and recommendation systems simply rank the presented items by their similarity to the user profile. In personalized search systems, the profile is fused with the query and applied for filtering and re-ranking of initial search results. Referring to any specific paper is difficult, since there are dozens of reported systems using
this approach: see  for a review.

TaskSieve was designed to assist users who perform exploratory searches reasonably often, i.e., it focuses on relatively experienced searchers up to the level of professional information analysts. These
users appreciate more powerful and sophisticated information access tools; but as we learned from our earlier work on adaptive filtering .

The flexibility in controlling the integration mode between queries and the task model also demonstrates its usefulness. First, we  observed subjects switching among the different modes in their
searches. Second, the searches with the half-half mode produced the best results. Third, the searches in query-only mode produced better results than the baseline, which indicates that the users really
mastered the preset manipulations and used the appropriate mode for different searches. Finally, it is clear that none of the modes significantly dominates all the searches. All of these indicate that it
really makes sense for TaskSieve to let users decide the best mode for their searches.