Mining Social Network for Personalized Email Prioritization

1
Mining Social Networkfor Personalized Email
Prioritization

• Language Techonology Institute
• School of Computer Science
• Carnegie Mellon University
• Shinjae Yoo, Yiming Yang, Frank Lin, and Il-Chul
  Moon

2
Outline

• Problem Description
• Approaches
• Experiments
• Contributions

3
Problem Description

• Email Overload is severe problem
• Identifying Importance of email will alleviate
  email overload
• Challenges
• No access to other peoples emails and labels
• Personalized labeling is time consuming
• The same message may have different priority
  labels for different recipients
• We want to leverage the sparse training data by
  using social network of each user

Sparse Training Data
4
Outline

• Problem Description
• Approaches
• Social Clustering
• Social Importance
• Semi-supervised Importance Propagation
• Experiments
• Conclusion and Future Work

5
Social Clustering Motivation

• Personal Email Inbox
• Lots of unlabeled emails
• No privacy issue
• Observations
• The sender can be important
• Some senders are not appeared in the training set
  at all or very few instances
• Need generalization of sender
• ? Lets find similar senders from social
  network

6
Social Clustering Contact Network

• Personal Contact Network
• G (V,E )
• All the network is constructed from personal inbox

Agent /Person
3
5
4
1
2
7
Social Clustering Newman Clustering

• Newman Clustering Algorithm Newman, 04
• Find social cliques or cohesive social groups
• Based on edge betweeness
• The number of shortest path that go through the
  edge / the total number of shortest path
• Drop edges from highest edge betweeness
• Hard clustering

9
1
4
4
4
4
4
2
3
5
6
1
1
Group A
Group B
8
Social Clustering Validations

• Clusters are coherent!

9
Social Clustering Feature Incorporation

• Extended Vector Space
• text social network
• combined
• The combined vector space is used as enriched
  feature set to the email prioritizer

10
Social Importance Motivations

• Social Importance
• A person in the center of a cluster might be more
  important than others
• Betweeness
• Edge betweeness for Newman Clustering
• Vertex betweeness
• The degree of communication bottleneck from
  social network
• Contact points among the network
• Might be important person
• We may try other kinds of social importance
  metrics too

11
Social Importance Metrics

• Metrics
• Degree (in, out, total) Wasserman and Faust, 94
• Clique Counts (ClqCnt) Wasserman and Faust, 94
• The number of clique sub-graphs which contain a
  node v
• Betweeness (BetCent) Freeman, 77
• HITS Authority (Authority) Kleinberg, 99
• ? the greatest Eigen value
• r the Eigen vector ? similar to PageRank scores
• Neighborhood Connectivity (Clustering
  Coefficient, ClustCoef) Boykin and
  Roychowdhury, 05
• measure the connectivity among the neighbor of a
  node v

12
Social Importance Validations

• Correlation coefficients with priority levels

13
SIP- Motivations

• Semi-supervised Importance Propagation (SIP)
• Can we propagate importance labels?
• Bi-partite graph, Labels only in Emails

Agent /Person
?
?
?
?
?
Emails
4
3
2
?
?
14
SIP- Email Network

• A Sender to Emails (N x M)
• BT Email to Recipients (M x N)
• xk kth importance labels for emails(M x 1)
• ykBxk (N x 1)

Agent /Person
?
?
?
?
?
Emails
14
4
3
2
?
?
15
SIP - Algorithm

• Problems of the above propagation
• may not be irreducible
• is insensitive to (not personalized)
• Apply Personalized PageRank with
• Normalize and column-wise normalize C
  C

16
Outline

• Problem Description
• Approaches
• Experiments
• Contributions

17
Experiments Data Collection

• Collected Data
• 25 subjects are recruited from Canegie Mellon
  University
• 7 users who submitted more than 200 emails
• 1 faculty, 2 staffs, 4 students

Training
Testing
time
18
Experiments Metrics

• Mean Absolute Error (MAE)
• 1.0 MAE means on average the prediction is
  deviated from the truth by one priority level
• MAE considers the difference among the errors
• It ranges from 0 to 4 when we use five importance
  level
• 1 vs. 5 and 4 vs. 5
• Micro-MAE
• Pooling the test instances from all users to
  obtain a joint test set
• Macro-MAE
• Compute each user MAE first and then take the
  average of per-user MAE

19
Experiments Setups

• Features four subsets
• Basic Feature (BF) from, to, cc, title, body
• Newman Clustering (NC)
• Social Importance (SI)
• Semi-supervised Importance Propagation (SIP)
• Ten times random shuffling among training data
• Linear SVM
• 10 Fold C.V. for parameter tuning
• Tuned regularization parameter 10-3.. 103

20
Experiments Results
21
Contributions

• The first study on personalized email
  prioritization
• Using statistical classification and clustering
• Based on fine-grained personal judgments with
  multiple users
• Enriched representation through personal Social
  Network
• Social Clustering
• Social Importance Estimation
• Semi-supervised Importance Propagation
• Fully personalized methodology
• Technical development and Evaluation