9. [scenario 1: route to work]
How do we
get
personalised
answers to
our queries?
10. [scenario 2: shopping]
get me to the closest
shop that accepts my
credit card where I
can buy some food
Who is
entitled
to upload
what
content
into the
system?
11. [future work]
Personalised Transport Information System
Trust-based recommender system
Reliable & Decentralised Data Collection
Trust-based access control
Analysis of the emerging communities of
travelers
Threat, Vulnerability, and Risk Analysis for
use of data
TU Eindhoven
Magma Srl
Cadzow Ltd.
13. Coping with unwanted content in MANet
Vision:
-A lot of content is created and
shared
Challenges:
-Users’ interest follows Zipf
distribution
Afra Mashhadi
14. Whom to forward to?
Avoid forwarding content to people who
are not interested:
How to know who is interested?
How to reach them?
16. Reasoning on mobility
Learn regularity of your contacts
Decide who should be the next hop
based on the probability of
meeting
And
the interest of the relayers
19. Co n t en t n o l o n ger c l a ssi f i ed i n a
Problem definition
h i er a r c h y wh i c h c a n b e n a vi ga t ed
i n o r d er t o f i n d i n t er est i n g
c o n t en t
• Content overload
• Personalization of content:
Social tagging behaviour
• Efficiently connect users with
relevant content within a huge
dataset
accuracy
coverage
19
20. Problem definition
•CiteULike social bookmarking website:
•Users have clearly defined interests: they bookmarked a small subset of papers using a
small subset of tags!
Standard
information
retrieval system:
?
Poor performances for queries that look for medium-to-low popularity content!
Accuracy for papers tagged only by a small subset of users
for tags used only by a small subset of users, due to the empty overlap
between tags
Coverage
20
21. Proposed solution: Social ranking
• Social ranking goal: efficiently connect users with
relevant content within a huge dataset
Accuracy: User similarity
Coverage: Tag expansion
Activity approach Dictionary approach
Similarity of users computed Similarity of tags computed
Similarity of tags computed on
on tags they used according to their semantic
papers they were associated to
relationship
21
22. Future works
Improve the efficiency (accuracy, coverage, scalability) of the
proposed technique
Performance still lowered in accuracy by noise caused by low
tagger users (more than 70% of low taggers!)
Apply clustering techniques to group
users into communities of interest/tags
into communities of topics
Apply clustering techniques to group
only power users (heavy taggers) into
communities and to infer the best fitting
group for each low tagger user
22
24. Improving Content Searching
in Social Tagging Systems
How to find exactly what I want?
How to locate relevant content?
How to discover important items
ranked based on my interests
Collaborative Filtering
How to improve it?
Claudio Weeraratne
25. Goals
•
Outperform the static model of the
system used by Collaborative
Filtering
•
Find more stable
algorithm/similarity measures
•
Point the focus on user's interest
and on the concept of community
26. Analysis
AIM: Increasing accuracy and coverage
Analyse users' similarity evolution
stability of interest between users
u u
4 4
u u u
u 1 1 2
u u
6
3 6
Analyse users' interest evolution
stability of interest per user
t t
4 4
u u
1 t 1
t t 6
6 3 t
2
27. Output
Get a interest-based view of the network
Clustering user by interest
If exists stability between users over time
Improve similarity method to achieve stability
If exists stability in users' interest
Find a time-adaptive similarity measure
If users' interest change over time
28. the future of mobisys seminars:
how can we bring our research together?