Loading presentation...

Present Remotely

Send the link below via email or IM

Copy

Present to your audience

Start remote presentation

  • Invited audience members will follow you as you navigate and present
  • People invited to a presentation do not need a Prezi account
  • This link expires 10 minutes after you close the presentation
  • A maximum of 30 users can follow your presentation
  • Learn more about this feature in our knowledge base article

Do you really want to delete this prezi?

Neither you, nor the coeditors you shared it with will be able to recover it again.

DeleteCancel

Make your likes visible on Facebook?

Connect your Facebook account to Prezi and let your likes appear on your timeline.
You can change this under Settings & Account at any time.

No, thanks

SIGIR 2012

Image Ranking Based on User Browsing Behavior
by

Michele Trevisiol

on 4 April 2014

Comments (0)

Please log in to add your comment.

Report abuse

Transcript of SIGIR 2012

Yahoo! Research Barcelona
Universitat Pompeu Fabra

Image Ranking Based on
User Browsing Behavior

Image Ranking on a Social Network
Michele Trevisiol
Luca Chiarandini

@ UPF & Yahoo! Research
@ Yahoo! Research
Luca Maria Aiello
Alejandro Jaimes

..for example, in
flick
r
importance of an image depends on
internal
and
external
factors
user that make actions
is also more likely to receive them
image that belongs to
many Flickr groups
is more likely to be discover
user's contacts
are
alerted
when the user upload new photos
image linked by many external websites
(e.g. twitter, blogs, ..)
more reachable from outside of Flickr
Investigate
internal
and
external factors
that affect image ranking.
Apply a
page-ranking approach
for image ranking
Exploiting Users Navigation Patterns for Image Ranking
represent aggregate
user navigation patterns as a graph
(BrowseGraph)
consider also the external URLs
apply
PageRank
on BrowseGraph
apply
BrowseRank
on BrowseGraph
Main Contributions
Contributions
we
compare five different

implicit and explicit
ranking methods
to a number of features

we introduce a
variation of BrowseRank and PageRank
in order to apply the
graph-based

methods on users navigation patterns




we
analyze the connectivity patterns
of a large BrowseGraph extracted from Flickr
Extraction of Sessions
Flick
r
sample of 2 months of data
10 million anonymous users (Aug-Oct 2011)
Anonymized user | timestamp | referrer URL | current URL | User-Agent
data filtering and cleaning
by
time
and by
external URL
about
309 million
pageviews
extraction of the BrowseGraph
pageview structure:
BrowseRank
[Yuting Liu et al., SIGIR 2008]
ranking algorithm
based on PageRank

relies on navigation graph and not on link graph

considers the
time that users spend
on each page
indicatior
of user's
interest
reset probability
:
number of sessions that start in j over the total number of sessions that contain j
stop probability
:
number of sessions that end in j over the total number of sessions that contain j
ranking methods
popularity, interestingness and diversity
Visual
inspection
1k images results
of
explicit
,
implicit
and
centrality-based
techniques
Favorites
absolute number of favorites assigned to a photo
explicit opinion
(only by Flickr users)
internal popularity
external popularity
collective attention
diversity
no. of comments
no. of Flickr groups
search engine results
Google PageRank
no. of photos linked
from external URL
no. of views
cumulative time spent
tag entropy
Comparison of Rankings
relief
internal factors
external factors
Flick
r
e.g.
search engine, blogs, social platforms, email, news, ..
session
exploiting
implicit
users behavior
considering both
internal
and
external
factors
sessions extraction
...
photo
user
group
photo
1.0
1.0
0.5
the weight depends on the number of non-entity nodes (
not photo
,
user
or
group
)
we
manually group external domains in 17 classes
, covering 99% of the total number of external URLs
one session:
top 10 most frequent external url classes
in our sample
internal popularity
collective attention
explicitly expresses the users' interests
implicitly expresses the users' interests
Comparison of Rankings
external popularity
diversity
BrowseRank
PageRank
Favorites
Time
View
avg. tagged
photos
number
of tags
unique
tags
avg. tags
per photo
entropy
(tags freq.)
0.73
0.75
0.53
0.80
0.83
7913
7129
4164
6192
6523
4347
3583
2936
2245
2113
7.93
7.39
5.98
6.20
7.14
11.23
10.57
10.81
9.31
7.14
artistic
high-quality landscapes or portraits
peculiar
or curious shots
part of specific
photo series
or serial events
major natural and social
events
Visual inspection of top 10 images
artistic
high-quality landscapes or portraits
major natural and social
events
part of specific
photo series
or serial events
peculiar
or curious shoots
3
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
1
2
2
2
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
4
4
4
4
external URL
considering links from outside the Flickr domain
top 10 results
observations on the rankings
favorites
photos well spread across Flickr
groups
(e.g. fear factory)
photos that receive attention from
active Flickr users
photos made by professionals
browserank and pagerank
photos with
higher semantic
variety
highlighted
events,

peculiar

and
fun topics
fully
implicit
minimize 'noise' of serial
images
views / time
easy computation
photos with very
low
diversity
of results
poorly overall
performance
Conclusions
performed
comparison among rankings on a large Flickr dataset
along several axes
internal and external
popularity
overall attention
that they attract from Web users
diversity in terms of
semantic
categories
visual
appearance

results show
that
rankings based on
explicit user feedback
behave better than simple implicit methods
favorites-based
method shows mainly artistic photos very active inside the network, but they are limited in variety and have low impact on the external Web.
centrality-based
methods, promote images that have attracted interest of external Web services and minimize the noise due to serial and periodically photos
Faves
BR
PR
Time
View
Our Extension
art
curious
series
events
Faves
BR
PR
Time
View
legend
networks are
like ecosystems
activity

within
the network
determines the
importance
of content
inside it
but also
external links
give content importance
share content
manage social relations
join groups
perform favorites
for example:
flick
r
(external factors)
(internal factors)
image belongs to various
Flickr group
meta-data information (as
tags
)
'cult image'
inside Flickr
Nightmare Fear Factory
332'700
views
300
favorites
0
comments
many
hyperlinks
outside Flickr
e.g. news articles, Blogs, ..
real world event
generic impact (audience)
Occupy Wall Street
276'500
views
150
favorites
50
comments
image belongs to various
Flickr group
meta-data information (as
tags
)
'cult image'
inside Flickr
Nightmare Fear Factory
332'700
views
300
favorites
0
comments
many
hyperlinks
outside Flickr
e.g. news articles, Blogs, ..
real world event
generic impact (audience)
Occupy Wall Street
276'500
views
150
favorites
50
comments
about the photo
[http://www.flickr.com/photos/katia_romanova/7756641062/in/photostream/]
taken: August 15,
2008
views:
2,837
faves:
17
comments:
17
in Flickr since May 2008
his faves: 101
contacts: 370
by Craig Maccubbin
about the photo
taken: August 10,
2012
views:
955
faves:
122
comments:
40
in Flickr since Feb 2012
her faves: 435
contacts: 61
by Katia Romanova
[http://www.flickr.com/photos/26569037@N04/2796382198/in/set-72157623141511442/]
Racing gravity
in-depth
analysis of the results
of several ranking algorithms
by common metrics
by metrics based on views
take into account the structure of Flickr in terms of
users' navigation paths
our assumption
the
navigation
patterns within a
social media platform
have a
strong impact
on the importance of the
content
nodes:

entities
viewed by users
edges:

transition
between entities
for each node j
transition probability from node i to node j
stop probability of i
(or dumping factor)
reset probability of j
absolute number of views of the photo page
View
Time
cumulative time spent by all the visitors of the photo page
PageRank
PageRank score on the photo page (apply on BrowseGraph)
computed
reset
and
stop
probabilities
BrowseRank
BrowseRank score on the photo page
computed
reset
and
stop
probabilities
View
and
Time
rank better tagged photos
BrowseRank
and
PageRank
rank better photos that
have more tags
BrowseRank
present photos with higher
tag diversity
x-axis: top N results ([1,1000] images)
y-axis: cumulative value of the features
x-axis: top N results ([1,1000] images)
y-axis: cumulative value of the features
Future Works
BrowseRank application to
query-dependent
image ranking

BrowseRank application to
other domains
like news services

deep analysis on the
impact
of the estimated
reset and stop probability
on a closer graph

user-study evaluation
(user perception, user satisfaction, ..)

analysis of the
impact
that
temporal patterns
have on the quality of the ranking
visual inspection
top 10 images partitioned in 4 high-level categories
Full transcript