An attention-based Ranking Model for social media
Post on 21-Jun-2015
2626 Views
Preview:
DESCRIPTION
Transcript
Nicolas Maisonneuve, research associate Centre for Advanced Learning Technologies, INSEAD
Application of a simple visual attention model to the communication overload problem
Tags: Information overload, Community, Social Media, Attention‐based Ranking model, visual attention model, Social computing
Context: European research project www.atgentive.com
Sept. 2007
Scenario 1: Online Community
Problem :
Is there a way to recommend me the most important messages ?
1) Avoiding uninteresting messages according my interests,
2) … except if it’s about an important issue in the community
Situation• Member of an active community• I’m overwhelmed by the unread messages• I only have 10 minutes to understand the highlights since my last login.
Scenario 2: Weblogs & Social Media
Situation
• I have subscribed to a lot of interesting blogs
• Now I’ m overloaded by too many posts
• I only have 10 minutes to read all my feeds
Same Question:
How rank them and read only the most important ones for me ?
Research problem
Question:In a rich information (and social) environment, How do I choose items (message, blog posting, .. ) due to my limited resources (e.g. time, or people) ?
Answer: in a rich information environment, information competes for the user’s attention (c.f Attention Economy)
I choose the most attractive items
Conception of an Attention‐based Ranking Model to select items
How does an item attract the user’s attention?
Similarity in vision• In a scene (visual rich environment), which area (item) will attract my attention? • how to predict where my attention will be guided? (Visual Search problem)
Approach • Use of a visual search model: “guided Search2.0” (J. Wolfe, 1994)• Turn visual signals into communication signals (Message Reader = eye to perceive the social activity)
How does an item attract the user’s attention?
Attention guiding the 2 types of features:• Top‐down features (User guidance)e.g. user searching a green object
• Bottom‐up features (Stimuli guidance) e.g. flashy object in a dark scene
Saliency (i.e. attractivity) of a signalThe saliency of a signal is computed as the (weighted) sum of the saliency for each attractive feature of the signal (e.g. color, size, intensity, motion,etc…)
The Visual attention model “Guided Search 2.0” ‐ 1/2
Process 1) For each attractive feature, the signals are computed into a Feature Map (i.e. their levels of saliency according to the feature)2) Mix of the feature Maps into a global Saliency Map
The Visual attention model Guided Search 2.0 ‐ 2/2
How does an item attract the user’s attention?
In your context of communication signals…
Question 1: What are the top‐down features (user’s interest profile) ? Question 2: What are the bottom‐up features? (i.e. attractive features without knowing the user’s intention)Question 3: How to compute a feature map?Question 4: how to compute the saliency map?
Question 1: What are the top‐down features? (User driven attention)
Top‐down features•Message’s Topic: focus on specific topics
•Message’s User: focus on specific users
VG Market IT Industry Research
User's vigilance profile in a IT Community (scenario 1)
userA userB userC
User's vigilance profile in a Social Network (Scenario3)
Simple Vigilance profile P For a given context K (e.g. a task to do) ,
P(k) = (C,W) with:‐ C = The set of concepts c (user, topic) I want to pay specially attention to in a signal‐W = their respective levels of vigilance wcfor the user
‐ + Limited capacity H ( ∑wc<H and wc>wmin )(I can’t want to pay attention to everything)
Vigilance feature map
Question 2: What are attractive bottom‐up features? (i.e. without knowing the user’s intention)
1) Exception (temporal/spatial)
‐ Unusual sender
‐ Unusual topic
‐ Unusual activity (cf 5)
2) About me
‐message audience focussed on me (mailing‐list vs. personal message)
3) User’s effort
‐ Type of Medium
(Text < Sound< Video)
4) Urgency
Lifecycle of the message (3 months<now)
‐ See also 5)
5) Other’s influence
‐ Collective attention (burst of activity)
‐ Explicit Attention asked (Subject: [URGENT]… )
Question 3: How to compute a feature map?
Computation of a bottom‐up feature mapE = the set of unread items e1, e2, .. , en • For each feature k , each item is computed by a function fk to give its saliency [0, 1] related to this feature•A feature map is Mk={fk (e1), fk (e2), .. , fk (en)}
Example: Simple Computation of the Burst of (reading) Activity featureDefinition: Burst = an abnormal high level of activity : Last week, in average, a message has been read 10 times, but the message A has been read 30 times.
Computation:r(e,∆t) = the number of readings of the message e during the interval ∆t, m = the mean of r(e, ∆t) for the set of messages read during ∆tfburst(e)= with 1<t1<t2 the bounds
0
1
m m*t1 m* t2 inf
Conclusion
Future work
• Partially implemented (collective activity observer, burst of Activity, Vigilance Profile)
• Need to be evaluated (how to configure the weight of each Feature in the global saliency map computation?)
Features of the Ranking Model• Based on a Visual Attention Model
Not only what the user expects ( bottom up feature)• Use of social factors to rank items. • Try to integrate the notions of limited capacity & vigilance•Adaptive to the context (possible change of the vigilance profile)
Thanks for your attention. ☺
Scenario 3: Traditional Communication
Problem
Is there a way to notify me on a new emails only if :
‐ it is related to my current task (e.g. message from UserB)
‐ Or it delivers unexpected but important information.
Situation :
• Growth of the user’s connectivity (globalization + internet)
• I’m currently collaborating on a specific task with userA and specially with userB.
• I receive a lot of emails that interrupt my work
• 4 hours spent managing emails per day by senior management (Guardian Unlimited Newspaper, 2007)
• Economic Impact of the interruption caused by email+online tools: $588 billion/year for the Us Economy (Basex Research, 2005)
top related