InterestMap

InterestMap
- Harvesting Social Network
Profiles for
Recommendation
Hugo Liu (MIT Media lab)
Pattie Maes (MIT Media lab)
Speaker: Huang, Yi-Ching
Outline
Introduction
Social Network Profiles
The InterestMap Approach
Recommendations by using InterestMap
Evaluation and Performance
Discussion
Introduction
Recommendation Systems become
more central to people’s lives
E-commerce site
Amazon.com, Ebay
Know new friends
Friendster, Orkut
Personal model v.s.User model
Catergoary-based representation
Example: Orkut
passions
Common interest
Social Network Profile
Domain-independent user models
Friendster, Orkut, MySpace
Distinguish passions from other category into
ontology identity descriptors
Items map into their respective ontology of
interest descriptors
InterestMap Approach
How to build InterestMap?
Steps:
Mine social network profiles
Exact out a normalized representation
Augment the normalized profile with
metadata to facilitate connection-making
Apply machine learning technique to learn
the semantic relatedness weights between
every pair of descriptors
Normalized Representation
Mine 100,000 personal profiles
“passions” and common interest categories
Use natural language procession
Newly segmented list contain casually-stated
keyphrase referring to different things
Normalized Representation
21,000 interest descriptor and 1,000 identity
descriptor
Use ODP(Open Directory Project), TV tome,
Wikipedia, All Music Guide …etc
Identity descriptor: use ODP
Increase the chances that the learning
algorithm will discover latent semantic
connection
Discount 0f 0.5
Map of Interests and Identities
Latent semantic analysis
Landauer, Foltz & Laham, 1998
Pointwise mutual information (PMI)
Network Ontology
Features:
Identity hubs: identity descriptor node
Behave as “hubs” in the network
Link to interest descriptor node
Appear frequency:
Identity descriptor : interest descriptor = 18 : 1
Taste clique
When cohesion of clique is strong, taste clique behave
much like a singular identity hub, in its impact on
network flow
Network Ontology
Recommendations
Use InterestMap
Finding recommendations by spreading
activation
Evaluation Features:
Impact that identity hubs and taste cliques in the
recommendations
Effect of using spreading activation rather than
PMI scores
Evaluation and Performance
Discussion
Tradeoff:
Fixed ontology versus open-ended input
Socially costly recommendation
Implicit and privacy --> no cost
Make sure for conscious rating --> some cost
Users list items in their profile --> great cost
Conclusion
Recommender systems provide some
suggestions of things to do and people
to meet
General personal model for people
behave “in the wild” on the Web
Using cultural and taste model to
recommendation