Why? - Collaborative Web Search: Who, What, Where, When, and Why - page 66

Some researchers have explored the factors affecting answer quality on Q&A sites.
Raban and Harper ( 2008 ) point out that a mixture of both intrinsic factors (e.g., perceived owner-
ship of information, gratitude) and extrinsic factors (e.g., reputation systems, monetary payments)
motivate Q&A site users to answer questions. Ackerman and Palen ( 1996 ) and Beenan et al. ( 2004 )
confirmed that intrinsic motivations, such as visibility of expertise and the feeling of making a unique
contribution, influence participation in such systems. Results regarding extrinsic motivators have
been more mixed - Hsieh and Counts ( 2009 ) found that market-based incentives did not increase
answer speed or high-quality answers, but Harper et al. ( 2008 ) found that fee-based sites produced
higher quality answers than free sites. These insights into how social motivations influence system
participation suggest that perhaps collaborative search systems might include such mechanisms,
such as a “search reputation system” reflecting either a user's general search skills or domain-specific
expertise; such reputation mechanisms could also be valuable in helping users identify potential
collaborators.

6.3 INTRODUCING POTENTIAL COLLABORATORS

People sometimes receive the assistance of others who share their tasks without the knowledge of
either the person providing or receiving the assistance. For example, the clicks Martha made following
her search for “asthma causes” may feed back into the ranking algorithm and result in a better result
list for someone else searching on asthma causes. Although this implicit collaboration does not fit
the definition of collaborative search, there is an opportunity for systems to automatically detect
when people have similar needs and introduce them serendipitously so that they can choose to form
a more active collaboration if they so desire. For example, question answering websites and message
boards often serve to create connections between people with shared interests, which could then
lead to future collaborative searches. Serendipitous groupings appear to be common in collaborative
search. In a diary study in which information workers recorded their collaborative search experiences
over a one-week period, Amershi and Morris ( 2009 ) found that the majority of such collaborations
(61.9%) were spontaneous, rather than planned.
Collaborative filtering is one way that data from similar people is identified for implicit use
in improving the search experience for an individual. As an example, Sugiyama et al. ( 2004 ) filled
in sparse user term-weight profiles by applying collaborative filtering techniques to provide term
weights based on those of users with similar profiles. Sun et al.'s CubeSVD approach ( Sunetal. ,
2005 ) used click-through data (represented as a user+query+URL triple) to generate personalized
Web rankings; they used collaborative filtering techniques to generate missing click-through triples,
thereby enhancing their technique's performance. Dou et al. ( 2007 ) compared several personalization
strategies and found that the use of click-through data and k-nearest neighbor collaborative filtering
was a promising approach. Almeida and Almeida ( 2004 ) used Bayesian algorithms to cluster users of
an online bookstore's search service into communities based on links clicked within the site, and they
found that the popularity of different links within different communities could be used to customize
search result rankings. VisSearch ( Lee and Y-J. , 2005 ) uses data mining to uncover patterns in users'