March 24th, 2011

wrendeceitInternet και Εφαρμογές Web

21 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

58 εμφανίσεις


Communicating Trust in Information Networks


Assoc

Prof. Tobias
Höllerer
, Dr.
John O’
Donovan

Dept. of Computer Science

University of California, Santa Barbara


Network Science CTA, INARC Midterm Review

March 24th, 2011




Objectives


Theory
of the human cognition/ and
perception involved in judgments of credibility
and trust


Understanding influences
of credibility and
trust on the subsequent decision
-
making


Modeling and representation of trust in
Information Network UIs

Research Focus


Conceptualizing Trust in Information Networks


(defining a scope)


What visual mechanisms are best for communicating
trust in an information network?


What is the role of interaction?



--
search and exploration



--
provision of trust information


How does a trust
-
based interface change the user
experience?

Overview

Social UIs

(
Smallworlds

/

PeerChooser

Initial

Semantic Links

(
DBpedia

Vis)

Core Visual/

Interaction
Platform

(
WiGis
)

Initial

Topic
-
based UIs

(
TopicNets
)


Models for
communicating
Trust

Combining
Social and
Semantic
Networks

Previously (I2.3)

Currently (T2.2)

Trust
-
based

Social UIs

(T
-
Smallworlds
)

Semantic Links

(
WiGiPedia
,

Semantics
-
based
Recommenders)

New Topic
-
based UI
incorporating

Social Trust

Larger

Scale

Iterative

User

Studies

Small
Worlds

Live Facebook application

apps.facebook.com/smallworlds


Recommends people and any
items from Facebook using ACF


Supports user interaction to
update profile info at
recommendation time


Supports user provided trust
gestures through click and drag


Makes the ACF algorithm
transparent and understandable.



Interactive, Trust
-
based Recommender
for Facebook Data

Small
Worlds

Live
Facebook

application

apps.facebook.com/smallworlds


Recommends people and any
items from
Facebook

using ACF


Supports user interaction to
update profile info at
recommendation time


Supports user provided trust
gestures through click and drag


Makes the ACF algorithm
transparent and understandable.



Interactive, Trust
-
based Recommender
for Facebook Data

User Evaluations

Combining Social and Semantic Data to find
Trusted Resources

Facebook /
Twitter

(
Social Recs
)

DBPedia
/Freebase

(Semantic
Recs)

Linked Data

Linking in the Semantic Web

Wigipedia

Demo

Ongoing Work: Semantically

Supported Recommendations

PARC Twitter Dataset. (Suh & Pirolli)


32,000 Twitter Profiles (bags of Tweets)


500 Mined Topics


3.27G corpus with 32K documents.


17K vocabulary


300M word tokens.


16

Topic Modeling


(
Porteous

et al., SIGKDD 2008)

topics are
distributions over
words

documents are a
mixture of topics

topic is a latent
variable

[
BAYESIAN INFERENCE
]
sampling
bayesian

prior distribution sample
monte_carlo

method

model

samples

posterior

markov_chain

inference

importance

gibbs

likelihood

parameter

bayes

mixture

mcmc

gaussian



[
DIGITAL LIBRARIES
]
digital library
libraries

access

collection

information

metadata

electronic

repository

repositories

catalog

archives

archive

providing

content

portal

resources …

Example
learned topics:

17

Hidden Markov Models in
Molecular Biology: New
Algorithms and Applications

Baldi, Chauvin, Hunkapiller, McClure


Hidden Markov Models (HMMs) can be applied to
several important problems in molecular biology. We
introduce a new convergent learning algorithm for
HMMs that, unlike the classical Baum
-
Welch
algorithm is smooth and can be applied on
-
line or in
batch mode, with or without the usual Viterbi most
likely path approximation. Left
-
right HMMs with
insertion and deletion states are then trained to
represent several protein families including
immunoglobulins and kinases. In all cases, the
models derived capture all the important statistical
properties of the families and can be used efficiently
in a number of important tasks such as multiple
alignment, motif detection, and classification.

[topic 10]

state hmm
markov sequence models
hidden states

probabilities sequences
parameters transition
probability training hmms

hybrid model likelihood
modeling


[topic 37]

genetic
structure chain protein
population region

algorithms human mouse
selection fitness proteins
search evolution

generation function
sequence sequences genes

[cluster 88]

model data models
time neural figure
state learning set

parameters
network
probability number
networks training
function system

algorithm hidden
markov


Multiple

Topics

One

Cluster


Topic
modeling
(LDA) better
than
clustering


(
Porteous
, Newman et al. 2008)

Mining Trust through Topic Modeling


LDA is a powerful
technique for
extracting salient
topics from text
collections.



IDEA:
Recommend
trustworthy info
feeds on Twitter
based on Topics



Expt:
Quantitative and
Qualitative
analysis

Iterative User
Evaluations



Topic
Nets

(INARC 2.3 2010 / TRUST 2.2 2011)


a) No Deformation

b) Single
-
Topic Deformation

c) All
-
Topic Deformation

Iterate

1:

Can we predict a user’s followers
based on analysis of the social network
.


2
:


Given user U's expertise in a set of topics
T
.

Can we
predict expertise
in topic
N


3
:


Given a user U's followers, can we predict their expertise
in terms
of Topics?
(also perform for followed users, and for
a combination
of both).

Which is the better
predictor and why
?


4
:


Comparative analysis of
user similarity in Twitter
using:
TillIT

Social
Net
analysis tool,
Topic
-
based
similarity and
Correlation based
similarity


(e.g. Pearson
,

Cosine
,

Jaccard
).


6
:

Compute trust in a given topic area

T
for a user
U
. This can be done by


1:
looking at the LDA association probability


2: looking at the
LDA associations to
Topic T of
the target
users Followers and Following
groups


Open
question:

how do we evaluate the computed "trust" score?



Possible techniques include:


a:

collect ground truth from real users.


b
:

leave
-
one
-
out analysis
and cross validation
techniques.

Ongoing and Future Work: Trust Modeling
Experiments using LDA Topics in Twitter Data


On
PARC’s

34,000 user Twitter dataset…

1:

UCSB Student at PARC for Summer ’11


2: Eye Tracker Study with
TopicNets

interface.


3: Phase 2 of the
TopicNets

redesign study and a publication of the findings.


4: Twitter Visualizations based on Trust and the Social Network (followers
followees
)


5: Improve / formalize model for communicating trust in user interfaces.


--
based on ongoing and iterative user studies


6: Evaluate combinations of social and semantic linkages for recommender systems.


--
discover a sweet spot?


7: Integration with ACT
-
R


--
towards development of a cognitive model of trust


--
ACT
-
R model of highly interactive interfaces













Path Ahead… (organized by distance)


Publications and Talks (UCSB)



TopicNets
: Visual Analysis of Large Text Corpora with Topic
Modeling,”ACM

Transactions on Intelligent Systems and
Technology, in the special issue on Intelligent Visual Interfaces
for Text
Analysis (2011, minor revisions)


VISSW workshop at IUI2011 on Eliciting Semantic Feedback
from
WikiPedia

user through interactive visualizations
.


2 submissions in preparation for
Visweek

2011.


Hollerer

Keynote talks at Int’l Symposium on Visual
Computing (ISVC 2010) and Winter Augmented Reality
Meeting (WARM 2011)


O’Donovan invited talks at UCI, PARC and UCSB CITS lecture
series


Thank You!


Scalable Interactive Visualization of >1M networked entities

An order of magnitude more scalable than the next best web
based graph tool

Native in all major
browsers, with no
plugins (flash, java
etc)

Overview and
Zoom Navigation

Seamless Transition between client (local) and server based
(remote) data models. Graph representations synchronized
through AJAX.

Visualize Remote
algorithms running
on local data in real
time

Feature Rich:

Layout, Clustering
and Interaction
algorithms

Rich search
functionalities, Node,
Edge, Group and
Shortest Path
Highlighting

Customizable
Semantic Framework.
Map any data field
-

(eg: Trust, Similarity)
to a Graph Dimension

WiGis Framework:

Research Challenges


What are the roles of both interface and interaction in
trust
-
based systems?


How does transparency introduced by the interface change
the user

s experience?


Does transparency and interaction (dynamic provision of
data) effect robustness and stability of a trust based
system?


How can Social and Semantic linkages be leveraged to build
a reliable, open and portable trust model for information
networks?

Inter
-
Task Coordination


T1.1
: Models and Metrics of Trust in Composite Networks
(
Mohapatra
)


T2.4: Network
Behavior Based Indicators
of
Trust
in
Composite
Social
and
Information Networks (
Adali
)



I2.2: Large
-
Scale Information Network Processing (Yan)


E1.2: Composite Network Modeling with Composite Graphs (
Basu
)


S3.1: Time
-
Stressed Decision
-
Making Under Uncertainty (Gray)