Recent developments in the area of social networking have lead to prominent users leaving facebook due to privacy concerns.
In order to really understand what motivated facebook to implement these controversial changes, you have to look at the future of recommender systems. I will introduce my current research in the areas of multi-source, cross-domain and privacy enabled user profiling and recommendation,
and show how it relates to current developments in the social networking space.
NO1 Qualified Best Black Magic Specialist Near Me Spiritual Healer Powerful L...
What your hairstyle says about your political preferences, and why you should care about the future of recommender systems
1. Digital Enterprise Research Institute www.deri.ie
What your hairstyle says about your political
preferences,
and why you should care about the future of
recommender systems
Benjamin Heitmann
Unit for Information Mining and Retrieval (UIMR)
Funded by Science Foundation Ireland under
Grant No. SFI/08/CE/I1380 (Líon-2)
Copyright 2009 Digital Enterprise Research Institute. All rights reserved.
Chapter
2. Why has Facebook been so controversial?
Digital Enterprise Research Institute www.deri.ie
Benjamin.Heitmann
slide 2 of 17
@deri.org
3. Why has Facebook been so controversial?
Digital Enterprise Research Institute www.deri.ie
new Facebook features
in April 2010
introduced new default
privacy settings
from private to public
by default
Benjamin.Heitmann
slide 2 of 17
@deri.org
4. Why has Facebook been so controversial?
Digital Enterprise Research Institute www.deri.ie
new Facebook features
in April 2010
introduced new default
privacy settings
from private to public
by default
huge backlash in the
media
Benjamin.Heitmann
slide 2 of 17
@deri.org
5. Why has Facebook been so controversial?
Digital Enterprise Research Institute www.deri.ie
new Facebook features
in April 2010
introduced new default
privacy settings
from private to public
by default
huge backlash in the
media
Result: simplified
privacy defaults
Benjamin.Heitmann
slide 2 of 17
@deri.org
6. Why has Facebook been so controversial?
Digital Enterprise Research Institute www.deri.ie
new Facebook features
in April 2010
introduced new default
privacy settings
from private to public
by default
huge backlash in the
media
Result: simplified
privacy defaults
Why change defaults in
the first place?
Advertisements are
recommendations, they
need user data!
Benjamin.Heitmann
slide 2 of 17
@deri.org
7. Recommendations have become a
commodity!
Digital Enterprise Research Institute www.deri.ie
Users expect “smart” web sites
Recommendations are part of such a user experience
Benjamin.Heitmann
slide 3 of 17
@deri.org
8. So why should you care?
Digital Enterprise Research Institute www.deri.ie
Recommendations
are a commodity:
they are required
for online products
Facebook Open Graph protocol
coming paradigm
shifts:
recommendations
outside the context of
a single site
much more
interesting / scary
recommendations
Benjamin.Heitmann
slide 4 of 17
@deri.org
9. The challenge
Digital Enterprise Research Institute www.deri.ie
Problem: How to compete
with existing systems?
provide relevant results
beyond your domain
share user profile data
beyond your site
encourage users to
trust you
Common aspect: vs.
multiple parties
eco-systems
architecture is new
required to define startup established
roles, standards and recommendation
interaction services
Benjamin.Heitmann
slide 5 of 17
@deri.org
10. Overview
Digital Enterprise Research Institute www.deri.ie
Background:
How are recommendations made?
Multi-Source recommendations:
Going beyond the context of a site
Privacy-enabled profiles:
Enabling the post-facebook future
Cross-domain recommendations:
So good, its almost scary
Benjamin.Heitmann
slide 6 of 17
@deri.org
11. Background: adaptive personalisation
Digital Enterprise Research Institute www.deri.ie
closed recommender system open recommender system
input data input data
recommendation recommendation
algorithm algorithm
background data integrated background data
external data sources source 1 source n
. . . .
Adaptive system: personalisation is based on an explicit user
model and profile data
3 components: Rec. algorithm uses background and input data
to make recommendations
Most existing rec. systems rely on collection of
in-house data and do not use external data
Benjamin.Heitmann
slide 7 of 17
@deri.org
12. Multi-source recommendations:
Digital Enterprise Research Institute www.deri.ie
First paradigm shift:
use background data
from external sources
link their items to your
users
extend the context of
the recommendations
beyond your site
Challenge: how to use
Linking the Facebook
Social Graph to structured data as
restaurants on Yelp
recommendation input
Solution:
use Linked Data
Benjamin.Heitmann
slide 8 of 17
@deri.org
13. Prototype: using Linked Data for multi-
source recommendations
Digital Enterprise Research Institute www.deri.ie
foaf:Person
myspace:topFriend
mo:MusicalArtist foaf:Person foaf:Person URIs
foaf:Document URIs
myspace (via DBTune) 0 1 0 1 1 0 1
0 0 1 0 1 0 1
foaf:interest
....
foaf:Document
foaf:Document
sioc:links_to
user-item matrix
FOAF vocabulary
sioct:WikiArticle
wikipedia (via SIOC exporter)
integrate with transform
apply collaborative
SPARQL CONSTRUCT from RDF graph
ltering algorithm on
query to matrix
user-item matrix
Use Linked Data to augment private data
this enables multi-source recommendations:
recommend new items through external background data
reduce sparsity by adding connections from ext. backg. data
provide recommendations for new users by using an external profile
Benjamin.Heitmann
slide 9 of 17
@deri.org
14. Evaluation of multi-source
recommendation results
Digital Enterprise Research Institute www.deri.ie
Smart Radio: first online
streaming and rec. radio
small: 190 users and 330 musicians binary cosine similarity
i1, i2: items e.g. musical artists
create links to MySpace artists
via DBpedia
external background data: 11000 relevant recommendations
users, 25000 new connections
collaborative filtering with
binary cosine similarity
evaluation compared to
Last.FM as “gold standard”
improve precision: 2% -> 14%
improve recall: 7% -> 33%
Benjamin.Heitmann
slide 10 of 17
@deri.org
15. Data sharing for recommendations
Digital Enterprise Research Institute www.deri.ie
2nd paradigm shift:
sharing user profile
data between sites
Reverse perspective:
?
users want to move
between social
networks!
data
Primary user
recommendations
sharing concern: privacy
Surprise: really hot
topic from this
perspective
Benjamin.Heitmann
slide 11 of 17
@deri.org
16. Facebook approach to privacy-enabled
user profiles
Digital Enterprise Research Institute www.deri.ie
The Facebook
approach:
express
preference centralised user
authentication
for user action
profile
data sharing for e.g.
web site
interaction recommendations
cross domain closed system
data sharing
if authorised no portability at
by user
all!
Challenge: open
alternative with
portability and
recommendations for privacy!
external site provided by
(at the same time)
facebook
Benjamin.Heitmann
slide 12 of 17
@deri.org
17. Alternative: architecture for private and
portable user profiles
Digital Enterprise Research Institute www.deri.ie
User profile:
Profile data expressed WebID
using RDF (FOAF+SIOC)
WebID provides identity
private key public key
(2 parts)
– private SSL Key in user
agent
– public SSL Key in FOAF user agent
FOAF Profile
profile
Roles: stored
user agents: manage user in
identities
profile storage service:
retrieves user profile
stores 1 or many profiles if user authorises it profile storage site
data consumer
data consumers: provide
services for users
Benjamin.Heitmann
slide 13 of 17
@deri.org
18. Communication pattern of the proposed
architecture
Digital Enterprise Research Institute www.deri.ie
WebID
private key public key
Storage URI
user agent
FOAF Profile
stored
in
profile storage site
data consumer
Benjamin.Heitmann
slide 14 of 17
@deri.org
19. Communication pattern of the proposed
architecture
Digital Enterprise Research Institute www.deri.ie
Scenario: restaurant
recommendation WebID
Assumption: user is logged into
Openbook
private key public key
Storage URI
user agent
FOAF Profile
stored
in
profile storage site
data consumer
Benjamin.Heitmann
slide 14 of 17
@deri.org
20. Communication pattern of the proposed
architecture
Digital Enterprise Research Institute www.deri.ie
Scenario: restaurant
recommendation WebID
Assumption: user is logged into
Openbook
private key public key
1. User requests nice restaurants Storage URI
from Chow
user agent
FOAF Profile
Any nice stored
restaurants? in
profile storage site
data consumer
Benjamin.Heitmann
slide 14 of 17
@deri.org
21. Communication pattern of the proposed
architecture
Digital Enterprise Research Institute www.deri.ie
Scenario: restaurant
recommendation WebID
Assumption: user is logged into
Openbook
private key public key
1. User requests nice restaurants Storage URI
from Chow
2. Chow gets profile storage via
user agent
Firefox FOAF Profile
Firefox stored
provides in
storage URI
profile storage site
data consumer
Benjamin.Heitmann
slide 14 of 17
@deri.org
22. Communication pattern of the proposed
architecture
Digital Enterprise Research Institute www.deri.ie
Scenario: restaurant
recommendation WebID
Assumption: user is logged into
Openbook
private key public key
1. User requests nice restaurants Storage URI
from Chow
2. Chow gets profile storage via
user agent
Firefox FOAF Profile
3. Chow redirects Firefox to
stored
Openbook for authorisation redirect to
openbook in
for authorisation
profile storage site
data consumer
Benjamin.Heitmann
slide 14 of 17
@deri.org
23. Communication pattern of the proposed
architecture
Digital Enterprise Research Institute www.deri.ie
Scenario: restaurant
recommendation WebID
Assumption: user is logged into
Openbook
private key public key
1. User requests nice restaurants Storage URI
from Chow
2. Chow gets profile storage via
user agent
Firefox FOAF Profile
3. Chow redirects Firefox to User authorises stored
Openbook for authorisation Openbook to in
show parts of
4. User authorises Openbook to profile to Chow
show some profile parts to Chow
profile storage site
data consumer
Benjamin.Heitmann
slide 14 of 17
@deri.org
24. Communication pattern of the proposed
architecture
Digital Enterprise Research Institute www.deri.ie
Scenario: restaurant
recommendation WebID
Assumption: user is logged into
Openbook
private key public key
1. User requests nice restaurants Storage URI
from Chow
2. Chow gets profile storage via
user agent
Firefox FOAF Profile
3. Chow redirects Firefox to
stored
Openbook for authorisation in
4. User authorises Openbook to
show some profile parts to Chow
5.Openbook redirects to Chow
redirect back to Chow profile storage site
data consumer
Benjamin.Heitmann
slide 14 of 17
@deri.org
25. Communication pattern of the proposed
architecture
Digital Enterprise Research Institute www.deri.ie
Scenario: restaurant
recommendation WebID
Assumption: user is logged into
Openbook
private key public key
1. User requests nice restaurants Storage URI
from Chow
2. Chow gets profile storage via
user agent
Firefox FOAF Profile
3. Chow redirects Firefox to
stored
Openbook for authorisation in
4. User authorises Openbook to Chow retrieves profile
show some profile parts to Chow parts now
5.Openbook redirects to Chow
profile storage site
6.Now Chow accesses parts of data consumer
profile data on openbook
Benjamin.Heitmann
slide 14 of 17
@deri.org
27. Cross-domain recommendations
Digital Enterprise Research Institute www.deri.ie
3rd paradigm shift:
Provide relevant results
beyond your domain
Requires user data
from multiple
domains
Benjamin.Heitmann
slide 15 of 17
@deri.org
28. Cross-domain recommendations
Digital Enterprise Research Institute www.deri.ie
3rd paradigm shift:
Provide relevant results
beyond your domain
Requires user data
from multiple
domains
Hunch.com shows
one solution: ask
your users
Benjamin.Heitmann
slide 15 of 17
@deri.org
29. Cross-domain recommendations
Digital Enterprise Research Institute www.deri.ie
3rd paradigm shift:
Provide relevant results
beyond your domain
Requires user data
from multiple
domains
Hunch.com shows
one solution: ask
your users
Benjamin.Heitmann
slide 15 of 17
@deri.org
30. Cross-domain recommendations
Digital Enterprise Research Institute www.deri.ie
3rd paradigm shift:
Provide relevant results
beyond your domain
Requires user data
from multiple
domains
Hunch.com shows
one solution: ask
your users
Benjamin.Heitmann
slide 15 of 17
@deri.org
31. Cross-domain recommendations
Digital Enterprise Research Institute www.deri.ie
3rd paradigm shift:
Provide relevant results
beyond your domain
Requires user data
from multiple
domains
Hunch.com shows
one solution: ask
your users
Benjamin.Heitmann
slide 15 of 17
@deri.org
32. Cross-domain recommendations
Digital Enterprise Research Institute www.deri.ie
3rd paradigm shift:
Provide relevant results
beyond your domain
Requires user data
from multiple
domains
Hunch.com shows
one solution: ask
your users
Benjamin.Heitmann
slide 15 of 17
@deri.org
33. Cross-domain recommendations
Digital Enterprise Research Institute www.deri.ie
3rd paradigm shift:
Provide relevant results
beyond your domain
Requires user data
from multiple
domains
Hunch.com shows
one solution: ask
your users
Benjamin.Heitmann
slide 15 of 17
@deri.org
34. Cross-domain recommendations
Digital Enterprise Research Institute www.deri.ie
3rd paradigm shift:
Provide relevant results
beyond your domain
Requires user data
from multiple
domains
Hunch.com shows
one solution: ask
your users
Alternative
solution:
automatically link
domains and
communities
Benjamin.Heitmann
slide 15 of 17
@deri.org
35. Future work: using Linked Data for
cross-domain recommendations
Digital Enterprise Research Institute www.deri.ie
Exploit the intrinsic
links between
John Cage Johnny Cash Elvis Metallica sources:
Myspace links between data
domain: music
from different sources
KyleButler TheTeacher
connections between
FOAF possible different domains
profile recommendations
identical users in
Dexter
Morgan different communities
Wikipedia
many domains
Country: City: Sport: Requires links between
Netherlands Amsterdam Sailing
data sources
(a.k.a. “The Linkage
Problem)
Benjamin.Heitmann
slide 16 of 17
@deri.org
36. Summary
Digital Enterprise Research Institute www.deri.ie
Recommendations have become a commodity
they are required for a good user experience
3 coming paradigm shifts:
Go beyond the context of one site
(multi-source recommendations)
Provide results beyond your primary domain
(cross-domain recommendations)
Enable eco-systems built around portable user profiles
(Privacy-enabled user profile portability)
Developing eco-systems with multiple parties
requires an architecture
(mutual agreement on roles, standards and
communication patterns)
Benjamin.Heitmann
slide 17 of 17
@deri.org
Editor's Notes
Start this slide by explaining the the major classes of rec. algo can be listed,
in increasing expense of the amount of knowledge which they require:
1.) CF is very cheap, only captures implicit knowledge
2.) CB requires automated feature extraction which is an open research problem for e.g. music or movies
3.) KB requires domain knowledge and annotations using this knowledge
4.) hybrid algorithms are used to balance the knowledge cost of algorithms, e.g. CF + something else
Start this slide by explaining the the major classes of rec. algo can be listed,
in increasing expense of the amount of knowledge which they require:
1.) CF is very cheap, only captures implicit knowledge
2.) CB requires automated feature extraction which is an open research problem for e.g. music or movies
3.) KB requires domain knowledge and annotations using this knowledge
4.) hybrid algorithms are used to balance the knowledge cost of algorithms, e.g. CF + something else
Start this slide by explaining the the major classes of rec. algo can be listed,
in increasing expense of the amount of knowledge which they require:
1.) CF is very cheap, only captures implicit knowledge
2.) CB requires automated feature extraction which is an open research problem for e.g. music or movies
3.) KB requires domain knowledge and annotations using this knowledge
4.) hybrid algorithms are used to balance the knowledge cost of algorithms, e.g. CF + something else
Start this slide by explaining the the major classes of rec. algo can be listed,
in increasing expense of the amount of knowledge which they require:
1.) CF is very cheap, only captures implicit knowledge
2.) CB requires automated feature extraction which is an open research problem for e.g. music or movies
3.) KB requires domain knowledge and annotations using this knowledge
4.) hybrid algorithms are used to balance the knowledge cost of algorithms, e.g. CF + something else
Start this slide by explaining the the major classes of rec. algo can be listed,
in increasing expense of the amount of knowledge which they require:
1.) CF is very cheap, only captures implicit knowledge
2.) CB requires automated feature extraction which is an open research problem for e.g. music or movies
3.) KB requires domain knowledge and annotations using this knowledge
4.) hybrid algorithms are used to balance the knowledge cost of algorithms, e.g. CF + something else
Emphasise the story: How can a start-up with very limited number of users and data,
go from their alpha phase to a competitive product? We have demonstrated that this is possible,
by augmenting the data from a small closed corpus recommender system.
Facebooks ecosystem is actually based on open standards, however it is not open.
It is a closed system.
Finish with:
Coming up with an alternative open system, and applying my research to such an open ecosystem,
will be a way to evaluate the contributions of my PhD.