Upload
ishmael-callahan
View
44
Download
0
Tags:
Embed Size (px)
DESCRIPTION
User Profiling in Ego-network : Co-profiling Attributes and Relationships. Rui Li, Chi Wang, Kevin Chen- Chuan Chang University of Illinois at Urbana-Champaign. User Profiling , which infers users’ attributes, is important for P ersonalized S ervices. User. Personalized Search. - PowerPoint PPT Presentation
Citation preview
The Database and Info. Systems Lab.University of Illinois at Urbana-Champaign
User Profiling in Ego-network:Co-profiling
Attributes and Relationships
Rui Li, Chi Wang, Kevin Chen-Chuan Chang
University of Illinois at Urbana-Champaign
User Profiling, which infers users’ attributes, is important for Personalized Services
2
and many others.
Personalized Search
Targeted Advertisement
Search Engines
Advertisers
Richard
User
College: UIUCLocation:
Champaign
3
User Profiling is crucial for Social Analysis– Ability to survey the world
Surveying people for behavior: How do college students like iPad vs. Galaxy? How do California age 50+ males like
ObamaCare?
Surveying behavior for people: What demographics of users like Samsung more
than Apple? What communities of people support ObamaCare?
Can we profile users’ missing attributes in social network?
4
Some users provide attributes in their online profiles
Some users’ attributes are missing
Employer: Yahoo! College:
Stanford
Employer: ? College:?
Employer: Yahoo! College: Berkeley
Employer: Twitter College: Berkeley
Employer: ? College:?
Employer: Twitter College: UIUC
Employee: ? College:?
Employer: Google College: UIUC
Employee: JP Morgan College: UIUC
Employer: ? College:?
5
Thus, we abstract our problem as profiling users' attributes based on friends’ attributes
Input: • a network G(V, E) ,• some users’
attributes
Output: • users’ attributes
Employer: Yahoo! College: Stanford
Employer: Yahoo! College: Berkeley
Employer: Twitter College: Berkeley
Employer: ? College:?
Employer: Twitter College: UIUC
Employer: ? College:?
Employer: JP Morgan College: UIUC
Employer: ? College:?
Employer: Yahoo!College: UIUC
While attributes may “propagate” across links—Links are very noisy.
6
Existing methods simply assume that two connected users share the same value for any attribute
Employer: Yahoo! College: Stanford
Employer: ? College:?
Employer: Yahoo! College: Berkeley
Employer: Twitter College: Berkeley
Employer: ? College:?
Employer: Twitter College: UIUC
Employer: ? College:?
Employer: JP Morgan College: UIUC
Employer: ? College:?
However, users connect to friends with different values for an attribute
Employer: Google College: UIUC
• About 11% friends share the employer and 18% friends share the college.
• Only 20% may have attributes.
Why noisy? Every link is for a (different) relationship!
7
Richard and Bob share the same employer, but may have different values
for other attributes.
Richard and Cindy share the same college, but may have different values for
other attributes.
Richard and Peter share the same interests, but may have different values
for other attributes.
Richard
BobColleagues
Cindy
Peter
College classmates
Club friends
Users have different types of relationships in real life.
8
On the other hand, Relationship Profiling is necessary by itself, and similarly challenged! Link: Why does a link happen?
Given a link, what friendship does it represent?
Circle: Who form what circles? Where are my circles? What does each circle represent?
Challenge: While links/circles depend on attributes to detect and to explain, attributes are often unknown.
9
Proposal: Co-profiling Attributes and Relationships Attributes– properties of nodes Relationships– properties of links Together, understanding both nodes and links.
Why together?
1. Necessity: Dependency on each other to decide.
2. Benefit: Useful to know both!
classmates Employer: Google
College: UIUC
Employer: Yahoo! College: Berkeley
colleaguesCollege: UIUC
Employer: Yahoo!
Missing Missing
10
But how?Observing how attributes and relationships
relate.
11
Insight: Correlation between attributes and connections through relationship
Discriminative Correlation Insight : Attributes and connections are discriminatively correlated via
a hidden factor -- relationship
To concretize our insight, we explore two dependencies based on a real-world user study.
• Attribute-Relationship Dependency: How users’ attributes are related to hidden relationship types?
• Connection-Relationship Dependency: How connections are related to hidden relationship types?
12
Observation #1: Attribute-Relationship Dependency
Friends do not share all attributes. What attributes they share depend on relationship.
The percentages of friends sharing the same value with the ego for different attributes overall of different relationship
types.
13
Observation #2: Connection-Relationship Dependency
Friends do not connect to all friends. What friends they connect to depend on relationship.
The average connections per user within and across three different relationships types
f3 =<1, 0, 0, 0, 1, 0, 0.1>
Specifically, we focus on co-profiling upon each user’s ego-network
14
Ego-network: a subnet that around an individual user.
Circle1: friends likely to share employee
Circle 2: friends likely to share college
Circle 3: friends likely to share other attribute
Employer: Yahoo! College: Stanford
Employer: ? College:?
Employer: Yahoo! College: Berkeley
Employee: Twitter College: Berkeley
Employer: ? College:?
Employer: Twitter College: UIUC
Employer: ?
College:?
Employer: Google College: UIUC
Employer: Yahoo
College: UIUC
Attribute Vectorf1 =<1, 0, 0, 1, 0, 0,
0.1>Circle
Assignmentx1=1
x3=1
Association Vector w1 =<1, 0, 0, 0, 0, 0, 0>w2 =<0, 1, 0, 0, 0, 0, 0>
f4 =<0, 1, 0, 0, 0, 1, 0 0.1>x4=2
Solution Overview: we realize co-profiling in an optimization framework
15
Unobserved Friends’ circles
Observed User Connections
Partially ObservedUser Attributes
Cost Function: capture the dependences between the variables based on the insight
Algorithm: finds the unknown variable that best satisfy the dependences
16
Cost Function: we design a cost function to model the dependencies between variables
Attribute-Relationship (circle) Dependency
Connection-Relationship Type (circle) Dependency
There are other formulas to model the dependencies.
However, the function can not be optimized directly, as there are both discrete and continuous
variables
17
Algorithm: we minimize the function via updating each group of variables
Update User Attribute Vectors F
Update User Circle Assignments X
Update Circle Association Vectors W
• Only propagate values from friends in the same circles
• Only propagate the attribute value associated with the circle
• Cosider both user’s attributes and connections
• Make association vector sparse
Experiment: we first collect real-world ego-networks to evaluate our data set
We conduct user studies to collect users’ attributes and relationship types (circles) from LinkedIn.
18
Ego Users Users Connections
175 19K 110K
We share the data online https://
wiki.engr.illinois.edu/display/forward/Dataset-EgoNetUIUC-LinkedinCrawl-Jan2014
• Most users are have three attributes • 8K connection are labeled
19
Experiment: we evaluate our algorithm on both attribute and relationship type profiling Attribute Profiling
APw: a classic collective classification approach, which profiles a node’s label using weighted votes from its neighbors.
APi: anther collective classification (semi-supervised learning) approach, which iteratively profiles nodes’ labels with APw.
APc: a state-of-art method, which profiles users’ attributers based on clustering network.
Relationship Type (circle) profiling RPa: profiles friends’ circles based on their attributes. RPn: profiles friends’ circles based on network structure RPan: profiles friends’ circles based on network and attributes, but assumes
attributes known.
20
CP is not only capable of profiling AP and RP and but also outperforms baselines for both
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7 Accuracy for College
APw APi APc CP0.48
0.5
0.52
0.54
0.56
0.58
0.6
0.62Accuracy for Employer
RPa RPn RPan CP0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
profiling college classmates
RPa RPn RPan CP0
0.1
0.2
0.3
0.4
0.5
0.6
profiling colleagues
Summary: we made the following contributions in this problem
We propose a co-profiling approach that jointly profiles users’ attributes and relationship types (circles) in ego networks.
We present the discriminative correlation insight to capture the correlation between attributes and social connections.
We conduct extensive experiments to evaluate our algorithms on two tasks based on real-world ego networks.
21
22
Thank You!