An Email and Meeting Assistant Using Graph Walks

1
An Email and Meeting AssistantUsing Graph Walks
Einat Minkov William W. Cohen
CEAS-2006
2
Documents and Links

• PageRank (Brin and Page, 98), HITS (Kleinberg,
  98)
• Co-training (Blum and Mitchell)

• Documents are not isolated objects they are
  connected to other documents via hyperlinks
• Document similarity/relatednessvia random graph
  walk

3
Structured Documents

• In structured data, documents are inter-connected
  via other common objects.
• Email and meeting entries are examples of
  structured datatext meta-data
• Represent email and meetings as a joint graph
• Derive extended similarity measures between graph
  objects using lazy graph walks.
• Show me recent relevant messages to this message
• What is the full name of Danny that is
  mentioned in this message?

Framework
Questions we can ask
4
Email as a Graph
Chris.germany_at_enron.com
Chris
alias
sent_from
sent_from_email
Mgermany_at_ch2m.com
sent_to_email
1.22.00
file1
On_date
sent_to
has_subj_term
Melissa Germany
has_term
work
where
yo
Im
you
5
Email as a Graph

• A directed graph
• A node carries an entity type
• An edge carries a relation type
• Edges are bi-directional (cyclic)
• Nodes inter-connect via linked entities.

6
Meetings

• Like Email messages, Meeting entries are
  structured.
• Share entities with Email
• Email and meetings can be naturally represented
  as a joint graph.

TIME
TEXT
PERSONS
7
The Joint Graph
nodex
Shared content
Social network
Timeline
8
Edge Weights

• Graph G - nodes x,y,z
• - node types T(x), T(y), T(z)
• - edge labels - parameters
• Edge weight x ? y
• Prob. Distribution

a. Pick an outgoing edge label
b. Pick node y uniformly
9
Graph Similarity

• Defined by lazy graph walks over k steps.
• Given

Stay probability
(larger values favor shorter paths)
A transition matrix
Initial node distribution
Output node distribution
We use this platform to perform SEARCH of
related items in the graph a query is initial
distribution Vq over nodes and a desired output
type Tout
10
Evaluation
Many tasks/ applications can be phrased as search
queries in this framework.
TASK I Find Meeting Attendees

• Given a meeting text date
• Retrieve a ranked list of relevant
  email-addresses (potential attendees)

TASK II Find Email Aliases
Given a persons name Retrieve a ranked list
of his/hers email-addresses
11
Methods
Corpus

• Baseline String matching Use distance metric
  (JARO-Winkler) Finds similar email-addresses
  to personal / project names mentioned.

• 346 email files (Meetings folder)
• 334 meeting entries (Palm)
• Both over the same time span (about 6 months)
• The joint graph includes 3,680 nodes

• Graph walk
• 3 Steps
• Uniform weights

12
Results Find Meeting Attendees
A. All email addresses

• 11-point precision-recall curve, averaged over
  13 examples

meeting
term
date
B. One address per person
file
e-address
13
Results Find Email Aliases
A. By first name

• 14 examples (2 to 5 email aliases each)

term
term
term
term
file
person
B. By full name
term
e-address
14
Summary

• A Joint representation of email and meetings
• Denser links
• Augments social network information
• Supports Meeting management applications
• Preliminary results are promising.
• Application of learning and more results for
  email-related tasks, available atContextual
  Search and Name Disambiguation in Email Using
  Graphs, Einat Minkov, William W. Cohen, Andrew
  Y. Ng in SIGIR 2006

15
Thanks! Questions?