An%20Online%20Social-Based%20Recommendations%20System

1
An Online Social-Based Recommendations System

• Danny Tarlow, Jeremy Handcock,
• Inmar Givoni, and Jorge Aranda
• CSC2231, December 2007

2
Online recommendations

• Same author, same genre
• Customers that bought this also bought
• Move towards more sophisticated algorithms
• Netflix challenge
• Attempt to use more information for the
  recommendations

Successful, but with room for improvement
3
Intuition of our approach

• Homophily
• We tend to get together with people that are like
  us
• We often have similar preferences
• New information available online
• What people like
• How they are connected socially
• Idea
• Link current machine learning recommendations
  technology with the social information now
  available

4
Problem and goals

• Goals
• Find out how to take advantage of social
  information for recommendation algorithms
• Build an application that implements our approach
• Test whether social data improve recommendations
• Our application
• Pulls preferences and social ties out of a
  community website
• Gives recommendations to users based on their
  preferences and ties
• Useful for recommendations in many current online
  social applications

4
5
Our subject

• We needed a website with publicly available
  information on preferences and social ties
• Boardgamegeek.com is the largest and most popular
  online community for boardgames and cardgames
  enthusiasts

gt32K games gt42K users, of which 30K have rated
games 1.3 million ratings gt128K social
(GeekBuddy) ties
5
6
Recommendations algorithm

• PMF Probabilistic Matrix Factorization
• Idea People are not that complex
• We can use the combination of a few descriptors
  for any of us (e.g., a strategy gamer who likes
  wacky themes)
• We also see how much each game fits to each
  descriptor
• We predict a user will like a game if it fits the
  descriptors of which he is made of
• Our algorithm finds these descriptors
  automatically
• Social information becomes part of our
  description for each user

6
7
Results and Application

• Using social information improves the performance
  of the PMF learning algorithm
• We plugged in the algorithm to our web application

7
8
Conclusion

• Contribution We developed an algorithm that
  takes into account social information
• Our algorithm increases the prediction accuracy
• We created a website to allow the gaming
  community to get better recommendations
• Open questions
• Incorporating preferences and social information
  from different websites?
• Identity management
• Does homophily really play a role here?

8