1. Problem & Motivation

• A central place to collect useful event information and to filter out unnecessary information is needed.
• To cope with various Collaborative Filtering recommendation (CF) related problems such as
• In order to tailor my information, tastes of unknown users who are calculated by the system to be similar to me has become the standard.
• Ad-hoc users
• Black-sheep users
• Data sparsity & Coldstart problem
• PITTCULT is a recommender based on trusted human network, utilizing human psychology to conform to the opinions of Friends

2. Background & Related Work

• Homophily - people with similar characteristics tend to be connected (Wellman, 2007) - on the Web
• Significantly similar interests with the friends who instant messaging often (Singla & Richardson, 2008)
• Similar tastes among friends and rating similarity is getting stronger in group (Groh & Ehmig, 2007)
• Friends¡¯ recommendation is better, more useful and trustworthy than system recommendation (Sinha & Swearingen, 2001)
• Advice seekers decide the value of suggested items according to the identity of the recommender (Bonhard & Sesse, 2006)
• Personal and subjective trust evaluation is more effective than CF and attenuates cold-start problem (Massa & Avesani, 2007)
• Profile and item trust based recommendation (O¡¯Donovan & Smyth, 2005)

3. Uniqueness of the Approach

• A real recommender application using trust ratings has rarely been explored.
• Epinions provides mechanisms for evaluating other people¡¯s overall opinions, regardless of the categories of the items, but does not provide for any personalized recommendation according to the evaluation.
• Ebay also lets its users rate trust values about others, but the rating values are only applied to the buying and selling transactions without recommendation.
• Moleskiing.it is a recommender application for ski mountaineers. Even an individual user can specify trustable users, the recommendation could be universal to every user (Avesani, et al., 2004).
• TrustMail implements ¡®Web-of-Trust¡¯ in a mailing application. Specifically, the system uses connected trust paths in a graph, thereby making it easier to propagate trust and, information (Golbeck & Hendler, 2004). However, it is hard to find the evidence that the application is in use
• There is also a movie system using a web of trust, but the trust value is calculated by the system, and the trust is for the recommendation agents, not for each user s(Bedi & Kaur, 2006).
• PITTCULT is the real trust-based recommender system which have been actively used

4. System Implementation

• Explicitly expressed trust value
  
  
  
  
• Implicitly expressed trust value
• When user saves an interesting event from another user¡¯s profile.
• When user answers invitation from friends positively.
• When user votes another user¡¯s review positively.
• Based on the accumulated user profile (own ratings and trust values for the others), recommendations are generated.
  
  

5. Results & Contributions

• This user study is to examine the usability of the system and user requirement for future implementation
• Number of participants was 8 users (7 students and one school staff)
  
  
• In free-text evaluation, users showed interest about negative evaluation.
• For future implementation, they require community support (M = 4.38), email function for sending recommendations (M = 4.38).
• They wanted to add their own event information and to search events by keyword.
• Planned Study
• Additional content-based recommendation as hybrid recommender will be implemented.
• Context-based recommendation based on users' input will be added - occasions, time or target audiences
• Reputation and information propagation will be implemented.
• Profile using user entered information - search keywords, tags and clicked facets - will be examined.
• Integration with existing social-web based system such as facebook

** This paper is also accepted to the Doctoral Symposium on the 2nd ACM International Conference on Recommender Systems (2008)