Trains of Thought: Generating Information Maps

Download Report

Transcript Trains of Thought: Generating Information Maps

Metro Maps of
Dafna Shahaf
Carlos Guestrin
Eric Horvitz
‘‘
The abundance of
books is a distraction
,,
Lucius Annaeus Seneca
4 BC – 65 AD
… and it does not get any better
• 129,864,880 Books (Google estimate)
• Research:
– PubMed: 19 million papers
(One paper added per minute!)
– Scopus: 40 million papers
Papers
Innovative
Papers
So, you want to understand
a research topic…
Now what?
Search Engines are Great
• But do not show how it all fits together
Timeline Systems
Research is not Linear
Metro Map
• A map is a set of lines of articles
• Each line follows a coherent narrative thread
• Temporal Dynamics + Structure
labor unions
Merkel
bailout
Germany
protests
junk
status
austerity
strike
Map Definition
• A map M is a pair (G, P) where
– G=(V,E) is a directed graph
– P is a set of paths in G (metro lines)
– Each e  E must belong to at least one metro line
labor unions
Merkel
bailout
Germany
protests
junk
status
austerity
strike
Game Plan
Objective
Algorithm
Does it
work?
Properties of a Good Map
1. Coherence
???
Coherence: Main Idea
Coherence is not a property of local interactions:
1
2
3
4
5
Greece
Debt default
Europe
Republican
Italy
Protest
Incoherent: Each pair
shares different words
Connecting the Dots
[S, Guestrin, KDD’10]
Coherence: Main Idea
A more-coherent chain:
1
2
3
4
5
Greece
Debt default
Austerity
Republican
Italy
Protest
Coherent: a small number of
words captures the story
Connecting the Dots
[S, Guestrin, KDD’10]
Words are too Simple
Bayesian
networks
Sensor
networks
1
Probability
Cost
Network
2
Social
networks
3
Using the Citation Graph
• Create a graph per word
– All papers mentioning the word
– Edge weight = strength of influence [El-Arini, Guestrin KDD‘11]
Network
3
2
1
6
4
7
5
8
Where did paper
8 get the idea?
9
Do papers 8 and
9 mean the
same thing?
Words are too Simple
Bayesian
networks
Sensor
networks
1
2
Social
networks
3
Probability
Cost
Network
Incoherent
Properties of a Good Map
1. Coherence
Is it enough?
Max-coherence Map
Query: Reinforcement Learning
Properties of a Good Map
1. Coherence
2. Coverage
Should cover diverse
topics important to
the user
Coverage: What to Cover?
• Perhaps words?
• Not enough:
1
SVM in oracle database 10g
Milenova et al
VLDB '05
2
Support Vector Machines in Relational Databases
Ruping
SVM '02
Similar Content
1
2
Different Impact
Citing Venues and Authors:
1
Affected more
authors/ venues
2
Very little
intersection
What to Cover?
• Instead of words…
• Cover papers
• A paper covers papers that
it had an impact on
• High-coverage map:
impact on a lot of the corpus
• Why descendants?
• Soft notion: [0,1]
p has High Impact on q if…
p
We use the algorithm of…
coherent
Many paths
(especially short)
r
Note that our protocol is
different from previous
work…
q
Formalize with coherent random
walks
Map Coverage
• Documents cover pieces of the corpus:
Corpus
Coverage
High-coverage, Coherent Map
Properties of a Good Map
1. Coherence
2. Coverage
3. Connectivity
Definition: Connectivity
• Experimented with formulations
• Users do not care about connection type
• Encourage connections between pairs of lines
Lines with No Intersection
Perceptrons
SVM
Face Detection
Optimizing Kernels
for SVM
SVM for Facial
Recognition
Solution: Reward lines that
had impact on each other
Tying it all Together:
Map Objective
• Coherence
– Either coherent or not: Constraint
Consider all coherent maps with
• Coverage
maximum possible coverage.
– Must have!
Find the most connected one.
• Connectivity
– Nice to have
Game Plan
Objective
Algorithm
Does it
work?
Approach Overview
Documents D
1. Coherence graph G
2. Coverage function f
f(
)=?
…
3. Increase
Connectivity
Coherence Graph: Main Idea
• Vertices correspond to short coherent chains
• Directed edges between chains which can be
conjoined and remain coherent
1
2
3
1 2 3 5 8 9
4
5
6
5
8
9
Finding High-Coverage Chains
• Paths correspond to coherent chains.
• Problem: find a path of length K maximizing
coverage of underlying articles
1
2
Cover(
3
1 2 3 4 5 6
)
?
4
5
6
5
8
>
9
Cover(
1 2 3 5 8 9
)
Reformulation
• Paths correspond to coherent chains.
• Problem: find a path of length K maximizing
coverage of underlying articles
a function of the nodes visited
• Submodular orienteering
– [ChekuriOrienteering
and Pal, 2005]
– Quasipolynomial time recursive greedy
– O(log OPT) approximation
Approach Overview: Recap
Documents D
1. Coherence graph G
2. Coverage function f
f(
)=?
…
Encodes all
coherent
chains as
graph paths
Submodular orienteering
[Chekuri & Pal, 2005]
Quasipoly time recursive
greedy3. Increase
O(log OPT)Connectivity
approximation
Example Map: Reinforcement Learning
multi-agent cooperative joint team
mdp states pomdp transition option
control motor robot skills arm
bandit regret dilemma exploration arm
q-learning bound optimal rmax mdp
Example Map Detail: SVM
Game Plan
Objective
Algorithm
Does it
work?
User Study
• Tricky!
– No double-blind, no within-subject
– Domain: understandable yet unfamiliar
– Reinforcement Learning (RL)
User Study
• 30 participants
• First-year grad student, Reinforcement
Learning project
• Update a survey paper from 1996
• Identify research directions + relevant papers
– Google Scholar
– Map and Google Scholar
– Baselines: Map, Wikipedia
Better
Results (in a nutshell)
Google
Us
Google
Us
Map users find better papers, and
cover more important areas
User Comments
Helpful
great starting point
noticed directions I didn't know about
… get a basic idea of what science is up to
why don't you draw words on edges?
Legend is confusing
hard to get an idea from paper title alone
Conclusions
• Formulated metrics characterizing good maps for
the scientific domain
• Efficient methods with theoretical guarantees
• User studies highlight the promise of the method
• Website on the way!
• Personalization
Thank you!