Why? - Collaborative Web Search: Who, What, Where, When, and Why - page 67

queries and browsing in order to generate recommendations for users with similar queries. Some
recommender systems, such as the movie recommender system PolyLens ( O'Conner et al. , 2001 ),
attempt to generate recommendation lists for groups of users.
Smyth, B. ( 2007 ) suggested that click-through data from users in the same “search commu-
nity” (e.g., a group of people who use a special-interest Web portal or who work at the same company)
could enhance search result lists. Smyth provided evidence for the existence of search communities
by showing that a group of employees from a single company had a higher query similarity threshold
than general Web users. Freyne and Smyth's I-SPY system ( Freyne and Smyth , 2006 ) expanded the
notion of search communities to include related communities, measuring intercommunity similarity
based on the degree to which communities' queries and result click through overlap. Mei and Church
( 2008 ) found that geographic location might serve as a reasonable proxy for community since they
observed that grouping users into classes based on the similarity of their IP addresses could improve
search results.
In all of the above cases, in addition to merely using data from other people to generate
recommendations, these implicit social-searching algorithms could provide a bridge to forming
more active collaborative partnerships. For example, group membership could be made explicit and
searchers could opt to actively work together on their shared task rather than merely passively con-
tributing their information. Analyzing query logs to determine when users may be simultaneously
engaged in related search tasks could be a way to help users form valuable collaborative relation-
ships ( Teevan et al. , 2009b ). Such systems for serendipitously suggesting collaborative search partners
could also benefit from infrastructures provided by social Web browsing tools, such as the Socia-
ble Web ( Donath and Robertson , 1994 ) or Community Bar ( http://www.communitybar.net/ ) ,
which allow users to chat with other people who are simultaneously viewing the same webpage.
When groups are explicitly brought together, there may also be the opportunity to use the
implicitly captured information they contribute to improve the search experience. Research on
personalizing search results ( Dou et al. , 2007 ; Smyth, B. , 2007 ; Teevan et al. , 2005 ) has found that
implicitly gathered information such as browser history, query history, and desktop information,
can be used to improve the ranking of search results on a per-user basis. Teevan et al. ( 2005 ) found
that the performance of the personalization algorithm they studied improved as more data became
available about the target user. This finding suggests that additional data from similar people may
be useful in enhancing personalization systems. A collaborative search system could evaluate various
metrics behind the scenes, including how similar previously viewed content is, to determine how
likely an explicit collaboration is to be of benefit to a particular implicit group were it to decide to
suggest it. Teevan et al. ( 2009b ) explored several ways to identify the value of group data, finding
that groupization (adapting personalization algorithms to incorporate data from a group of users)
could yield improved relevance in search result rankings. The improvement yielded by groupization
was dependent on aspects of the users' relationships to each other and on their current task. By
automatically identifying situations in which methods such as groupization would yield substantial