The Social Hourglass:
Enabling Socially-aware
Applications and Services
Adriana Iamnitchi
University of South Florida
anda@cse.usf.edu
Much Social Information Available
• Connects people through relationships
– Object centric: use of same objects
– Person centric: declared relationships or co-participation
in events, groups, etc.
Mining Social Data
•
•
•
•
•
•
Spam filtering
Sybil identification
Personalized search
Target marketing
Medical emergency notifications
…
Current Approach: Vertically Integrated
Socially-aware Applications
Challenges with Current Approach
• Application-limited collection and use of
social information
– High bootstrap cost
– Limited (potentially inaccurate) information. E.g.,
Information from online social networks
• Hidden incentives to have many “friends”
• All relationships equal
• Symmetric relationships
• Newer proposals to merge different sources
of social (and sensor) information for one app
– Specifically targeting context awareness
5
Motivating Application: CallCensor
6
Motivating Application: Sofa Surfer
7
Motivating Application:
Data Placement
8
Proposal: An Infrastructure for
Social Computing
Sofa Surfer
Roommate Finder
CallCensor
…
Objective
An infrastructure that:
• Can fuse information from various sources
• Allow user to control own information
– What is collected
– Where it is stored
– Who can access it
• Provide social knowledge to a variety of
applications:
– Social inferences (may be non-trivial)
10
Outline
•
•
•
•
•
Motivation
The Social Hourglass architecture
Social Sensors (work in progress)
Personal Aggregator (some ideas)
Social Knowledge Service: Prometheus (Kourtellis et al,
Middleware 2010)
– Data Management
– API for social inferences
– Experimental evaluation (on PlanetLab)
• Summary
11
The Social Hourglass Architecture
Applications
Sofa
Surfer
Roommate
Finder
CallCensor
Applications make use
Social Inference
API
social kno
Social knowledge servic
storing social data. Stor
Management
applications
A1
A2
Social Data
Personal Aggregators
Personal aggregators c
A3
sens
S11
S21
S22
S32
S33
S43
SocialSensors
sensors analyze
Social
socia
Social signals
Social Signals
12
Social Sensors
Consume existing social signals
• Location
• Collocation
• Schedule (e.g., Google calendar)
• Mobile phone activity (calls, sms)
• Online social network
interactions
• Email
• Personal relations (family)
• Shared content
• Shared interest (e.g., CiteULike)
• …
13
Social Sensors
• Report on behalf of ego:
– Alter, the person ego is interacting with
– An activity tag: e.g., “outdoors”, “dining”
• Based on content, location, predefined labels, etc.
– A weight: e.g., 0.15
• Run on ego’s mobile devices, desktop, or on web
• Processes user interactions
– To reduce noise
– To distinguish between routine and meaningful interactions
14
Social Sensors: Challenges
• Identifying activity tags:
– Mine text for keywords (emails, sms, blogs, etc)
– Reverse geo-coding to find where (co)located
– Predefined labels or dictionary and ontologies
• Quantifying interactions (assigning weights):
– Frequency, duration, time in-between interactions
– Familiar strangers versus active social interaction
15
Work in Progress: Social Sensor
for Gaming Interactions
• Variability in playing habits
• Variability in playing skills
• Time patterns
Aggregators
• Act as the user’s personal assistant
• Runs on trusted device (cell phone)
• Responsible for
– Managing passwords for various applications
– Personalization
– Identity management
Carol
Bob's
Identity
Manager
Carol
carol@work.com
User1
User2
carol@home.com
Alice's
Identity
Manager
@carol_hates_alice
The Social Hourglass Architecture
Applications
Sofa
Surfer
Roommate
Finder
CallCensor
Applications make use
Social Inference
API
social kno
Social knowledge servic
storing social data. Stor
Management
applications
A1
A2
Social Data
Personal Aggregators
Personal aggregators c
A3
sens
S11
S21
S22
S32
S33
S43
SocialSensors
sensors analyze
Social
socia
Social signals
Social Signals
18
Social Graph
19
Prometheus
• Peer-to-peer architecture
– Users contribute resources (peers)
– Fundamental change from typical peer-to-peer networks:
not every user has its peer
• Input: Social information collected from different
social sensors (reported via aggregators)
• Output: Social information made available to
applications and services
– Information made available subject to user policies
20
Distributed Social Graph
21
Prometheus Architecture
23
Architecture Details
• Users have a unique user ID
• Select trusted peer group based on offline
social trust with peer owners
• A user’s trusted peers communicate via Scribe
• Only the user’s trusted peers can decrypt
user’s social data and thus perform social
inference functions
24
Social Data Protection
• 2 sets of public/private keys
– User’s
– User’s trusted peer group
• Social sensors submit data encrypted with the group’s
public key and signed with the user’s private key
– Access to user’s private key only on user’s devices
– Data stored in the Pastry overlay
• Only trusted peers can decrypt and authenticate data
25
Social Inference Functions
The social graph management service exports an API that
implement social inferences
26
API for Applications:
Social Inference Functions
• 5 basic social inference functions:
• relation_test (ego, alter, ɑ, w)
• top_relations (ego, ɑ, n)
• neighborhood (ego, ɑ, w, radius)
• proximity (ego, ɑ, w, radius, distance)
• social_strength (ego, alter)
• More complex functions can be built
27
Social Strength
•
•
•
•
Quantifies strength between ego and alter
Result normalized to consider overall activity
Search all paths of maximum 2 social hops
One approach to quantify social strength.
Others are certainly possible.
28
Lessons from Experiments on
PlanetLab
• Social-based mapping of users onto peers
leads to significant performance gains:
– More than 15% of requests finish faster
– An order of magnitude fewer messages
• Reasonable latency
– Code significantly improved since publication in
Middleware 2010
29
Experimental Results:
Neighborhood Requests
10 users per peer
50 users per peer
Prometheus: User-Controlled P2P Social Data Management for Socially-Aware
Applications, Nicolas Kourtellis, Joshua Finnis, Paul Anderson, Jeremy Blackburn,
Cristian Borcea, Adriana Iamnitchi. 11th International Middleware Conference, Bangalore,
30
India, November 2010.
Real Social Traces:
NJIT Social Graph
100 randomly selected students
from NJIT given Bluetoothenabled phones that report
their collocation
• Data recorded
– Collocation with two
thresholds (45 and 90
minutes)
– Facebook friendships
• Sparse graph (commuters)
31
CallCensor
• CallCensor implemented on Android
– Cell phone silenced, rings or vibrates depending on the
social context and relationship with caller
– Relationship with caller:
• Social strength > threshold: allow call
• Caller directly connected by work
• Caller connected by work and ≤ 2 hops away
• Real social data from 100 users stored on 3 nodes
from PlanetLab
• Real time performance constraints
32
Lessons from CallCensor
Experiments
33
Resilience to (Social) Attacks
• Vulnerability to
malicious users
mitigated by directed,
multi-edged, weighted
social graph
• Vulnerability to
malicious peers related
to social graph
distribution
• Peers gain the
properties of the social
graph they represent
Summary
• The social hourglass architecture
• Prometheus: a decentralized service that enables
socially-aware applications and services by collecting,
managing and exposing social knowledge, subject to
user-specified privacy policies.
• Unique contributions:
–
–
–
–
Social graph representation
Aggregated social data
Social inference functions
Socially-aware design
35
Much Work to Be Done
• Developing social sensors
• Aggregator:
– proof of concept implementation
– Performance
• Evaluating benefits of social knowledge in
system design
• Socially-aware applications
• Query language for social inferences
• Privacy protection
36
More Information
• The Social Hourglass: an Infrastructure for Socially-aware
Applications and Services, Iamnitchi et al., IEEE Internet
Computing, May/June 2012
• Prometheus: User-Controlled P2P Social Data Management
for Socially-Aware Applications, Kourtellis et al., Middleware
2010
• Vulnerability in Socially-Informed Peer-to-Peer System, Jeremy
Blackburn, Nicolas Kourtellis, and Adriana Iamnitchi. Fourth
Workshop on Social Network Systems (SNS 2011)
http://www.cse.usf.edu/~anda
anda@cse.usf.edu
37
Acknowledgements
• My team of talented graduate students and
alumni:
• US National Science Foundation grants CNS0831785 and CNS-0952420
38
Thank you!
39
Neighborhood Inference
NUMBER OF USERS RETURNED
100
90
CL.90 and FB
80
CL.45 and FB
70
CL.90
60
FB
CL.90 or FB
50
40
CL.45
CL.45 or FB
30
20
10
0
1
2
3
4
5
6
SOCIAL HOPS FROM SOURCE
40
Social Strength Inference
CL.45 or FB
FB
CL.90
CL.90 and FB
1.0
0.9
SOCS VALUE
0.8
CL.45
CL.90 or FB
CL.45 and FB
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
1
2
SOCIAL HOPS FROM SOURCE
41
A Distributed System
42
42
Or a Distributed System
43
43
An Example: Interest Sharing
“Yellow Submarine”
“Les Bonbons”
“No 24 in B minor, BWV 869”
“Les Bonbons”
“Yellow Submarine”
“Wood Is a Pleasant
Thing to Think About”
“Wood Is a Pleasant
Thing to Think About”
The interest-sharing graph GmT(V, E):


V is set of users active during interval T
An edge in E connects users who share at least m file
requests within T
44
Small Worlds
Avg. path length ratio (log scale) .
10.0
Food
web
Power
grid
LANL
coauthors
Film
actors
Web
1.0
Internet
Word
co-occurrences
0.1
1
10
100
1000
10000
Clustering coefficient ratio (log scale)
D. J. Watts and S. H. Strogatz, Collective dynamics of small-world networks. Nature, 393:440-442, 1998
45
R. Albert and A.-L. Barabási, Statistical mechanics of complex networks, R. Modern Physics 74, 47 (2002).
Web Interest-Sharing Graphs
Avg. path length ratio (log scale) .
10.0
300s,
1file
Web data-sharing graph
Other small-world graphs
1800s,
10file
7200s,
50files
1.0
1800s,
100files
3600s,
50files
0.1
1
10
100
1000
10000
Clustering coefficient ratio (log scale)
46
DØ Interest-Sharing Graphs
Avg. path length ratio (log scale) .
10.0
Web data-sharing graph
D0 data-sharing graph
Other small-world graphs
1.0
28 days,
1 file
7days,
1file
0.1
1
10
100
1000
10000
Clustering coefficient ratio (log scale)
47
KaZaA Interest-Sharing Graphs
Avg. path length ratio (log scale) .
10.0
Web data-sharing graph
D0 data-sharing graph
Other small-world graphs
Kazaa data-sharing graph
2 hours
1 file
1.0
4h
2 files
28 days 12h
1 file
4 files
1 day
2 files
7day,
1file
0.1
1
10
100
1000
10000
Clustering coefficient ratio (log scale)
48
Proactive Information Dissemination
100
D0
Except largest cluster
90
Total hit rate
80
70
60
50
40
30
20
10
0
Web
3 days
Except largest cluster
Total hit rate
100
90
80
70
60
50
40
30
20
10
0
2 min
5 min
15 min
30 min
7 days
Kazaa
10 days 14 days 21 days 28 days
Except largest cluster
Total hit rate
100
90
80
70
60
50
40
30
20
10
0
1 hour
4 hours
8 hours
49