EVENT IDENTIFICATION IN SOCIAL MEDIA

1
EVENT IDENTIFICATION IN SOCIAL MEDIA

• Hila Becker, Luis Gravano Mor Naaman
• Columbia University Rutgers University

2
Social Media Sites Host Many Event Documents

• Event something that occurs at a certain time
  in a certain place Yang et al. 99
• Popular, widely known eventsPresidential
  Inauguration, Thanksgiving Day Parade
• Smaller events, without traditional news
  coverageLocal food drive, street fair

Photo-sharing Flickr Video-sharing YouTube
Social networking Facebook
Social media documents for All Points West
festival, Liberty State Park, New Jersey, 8/8/08
3
Identifying Events and Associated Social Media
Documents

• Applications
• Event search and browsing
• Local search

• General approach group similar documents via
  clusteringEach cluster corresponds to one event
  and its associated social media documents

4
Event Identification Challenges

• Uneven data quality
• Missing, short, uninformative text
• but revealing structured context available
  tags, date/time, geo-coordinates
• Scalability
• Dynamic data stream of event information
• Unknown number of events
• Necessary for many clustering algorithms
• Difficult to estimate

5
Clustering Social Media Documents

• Social media document representation
• Social media document similarity
• Social media document clustering
• Clustering task definition
• Ensemble algorithm combining multiple clustering
  results
• Preliminary evaluation

6
Social Media Document Representation
Title
Description
Tags
Date/Time
Location
All-Text
7
Social Media Document Similarity
Title

• Text tf-idf weights, cosine similarity

Title
Description
A
A
A
B
B
B
Description

• Time proximity in minutes

Tags
Tags
Date/Time-Keywords
time
Date/Time
Location-Keywords

• Location geo-coordinate proximity

Date/Time-Proximity
Location
Location-Proximity
All-Text
All-Text
8
Social Media Document Clustering Framework
Social media documents
Document feature representation
Event clusters
9
Clustering Ensemble Algorithm
Ctitle
Ensemble clustering solution
Consensus Function combine ensemble similarities
Wtitle
f(C,W)
Wtags
Ctags
Wtime
Ctime
Learned in a training step
10
Clustering Measuring Quality

• Homogeneous clusters

?

• Complete clusters

?

• Metric Normalized Mutual Information
  (NMI)Shared information between clustering
  solution and ground truth

11
Experimental Setup

• Data gt270K Flickr photos
• Event labels from Yahoo!s upcoming event
  database
• Split into 3 parts for training/validation/testing
  
• Clusterers single pass algorithm with centroid
  similarity
• Weighing scheme Normalized Mutual Information
  (NMI) scores on validation set
• Consensus function weighted average of
  clusterers binary predictions
• Final prediction step single pass clustering
  algorithm

12
Preliminary Evaluation Results

• Individual clusterer performance
• Highest NMI Tags, All-Text
• Lowest NMI Description, Title
• Ensemble performance, compared against all
  individual clusterers
• Highest overall performance in terms of NMI
• More homogenous clusters each event is spread
  over fewer clusters

Details in paper
13
Future Work Alternative Choices

• Document similarity metric
• Ensemble approach
• Weight assignment
• Choice of clusterers
• Train a classifier to predict document similarity
• Features correspond to similarity scores
• All-text, title, tags, time, location, etc.
• Numeric values in 0,1
• State-of-the-art classifiers SVM, Logistic
  Regression,

14
Future Work Alternative Choices

• Final clustering step
• Apply graph partitioning algorithms
• Requires estimating the number of clusters
• Evaluation metrics beyond NMI
• Datasets
• Flickr LastFM, YouTube
• Exploit social network connections

15
Conclusions

• Identified events and their corresponding social
  media documents
• Proposed a clustering solution
• Leveraged different representations of social
  media documents
• Employed various social media similarity metrics
• Developed a weighted ensemble clustering approach
• Reported preliminary results of our event
  identification approach on a large-scale dataset
  of Flickr photographs