Event Detection with Common User Interests

1
Event Detection with Common User Interests

• Hu Meishan, Sun Aixin, Lim Ee-Peng
• School of Computer Engineering, Nanyang
  Technological University
• School of Information Systems, Singapore
  Management University
• WIDM08

2
Outline

• Introduction
• Background events and user interests
• Motivation
• Event detection with common user interests
• Problem definition
• Proposed solution
• Query profile and its properties
• Online event detection
• Experimental evaluation
• Future work
• Conclusions

3
Event detection the traditional approach

• Task
• Given a stream of news articles, group them
  according to the events they describe.
• Drawbacks
• Many detected events may not be interesting to
  users
• Events that are interested to users but not
  heavily reported in news are not detected.

4
User interests and events

• User-created content reflect their interests
• many bloggers discuss their interested topics in
  their posts
• many users search for documents about their
  interested topics by submitting keyword queries.
• When an event is happening, we often observe
• a large number of blog posts discussing the event
  published
• a surge in the number of event related queries
• Popular queries are often event-related

5
An example of popular queries and the related
event
Mentions by Day number of posts mentioning
benazir bhutto per day in the past 30 days.

• Event
• Benazir Bhutto assassinated on 27 Dec 07

Source http//www.technorati.com/pop (top
searches captured on 28 Dec 07)
6
Motivations

• User-interested events can be detected by
  utilizing the user-created content, e.g.,
  queries, blog posts, etc.
• Challenges
• Not all popular queries are event related.
• Multiple queries might be related to the same
  event.
• Not all documents in the stream are worth
  processing in the event detection.

7
Outline

• Introduction
• Background events and user interests
• Motivation
• Event detection with common user interests
• Problem definition
• Proposed solution
• Query profile and its properties
• Online event detection
• Experimental evaluation
• Future work
• Conclusions

8
A new framework on top of existing systems

• A 4-step approach
• Popular query identification
• Query profile construction
• Event-related profile identification
• Online event detection

9
Temporal query profile and properties
10
Characteristic of event-related query profiles

• If a query q issued at time t is related to some
  event, there is likely a large number of
  documents describing the event matching q
  published within a short period before t.
• A event-related query often demonstrated
• a large number of documents matching it at the
  search time.
• a short time-span among documents in the
  constructed query profile.

11
Characteristic of profiles related the same event

• If query q and q are related to the same event
  at time t, there likely is a large number of
  documents describing the event matching both q
  and q published close to t.
• Two query profiles can be grouped into the same
  event if they are similar in content.

12
Characteristic of profiles describing event
evolution

• If a query q is related to an event that lasts
  for some time, the documents matching q for two
  searches at time t1 and t2, both within the
  period of the event, are likely to describe the
  evolution of the event.
• Not only similarity but also novelty between
  query profiles determines whether a query profile
  should be included into an event.

13
Online incremental clustering illustrated
14
Data and statistics

• 1 query stream
• the most popular 15 queries published by a blog
  search engine requested every 3 hours from
  2006-11-08 1AM to 2008-03-31 10PM
• 2 independent document streams
• TR blog posts traced by Technorati,
  http//www.technorati.com
• GN news articles traced by Google News,
  http//news.google.com

15
Accuracy of the event detections

• True event If the event is recorded in a
  Wikipedia article and the time of the recorded
  event is within a short period of the detected
  duration.
• Segment event If the detected event is wrongly
  split from a true event to which it should
  belong.
• Mixed event If the detected event contains
  queries and documents from two or more events
  recorded in Wikipedia.
• Unknown event If we cannot locate an entry
  recording the event in Wikipedia.

16
Outline

• Introduction
• Background events and user interests
• Motivation
• Event detection with common user interests
• Problem definition
• Proposed solution
• Query profile and its properties
• Online event detection
• Experimental evalution
• Future work
• Conclusions

17
Future work

• In the current framework, events are detected
  from one query stream and one document stream,
  however, it is possible to detect events from
  multiple query streams and multiple document
  streams.
• E.g., to associate the query profile constructed
  in blog data stream to that constructed in news
  data stream.
• A novel interface is in demand for browsing and
  searching the detected events.

18
Outline

• Introduction
• Background events and user interests
• Motivation
• Event detection with common user interests
• Problem definition
• Proposed solution
• Query profile and its properties
• Online event detection
• Experimental evaluation
• Future work
• Conclusions

19
Conclusions

• Motivated by the close relation between
  user-created content and real world events, we
  defined the problem of detecting events of common
  user interests.
• To address the problem, we proposed
• a framework that extents traditional event
  detection approach by seamlessly integrating the
  stream of documents and the stream of popular
  queries to form a stream of query profiles.
• a notion of query-profile and its properties that
  can facilitate the process of event detection.
• We use real world data in experiments and
  achieved high detection accuracy.

20

• Thanks!

21
Appendix statistic for profile filtering
22
Appendix parameters used in the detection