Combining Proactive and Reactive Predictions for Data Streams

1
Combining Proactive and Reactive Predictions for
Data Streams

• Ying Yang, Xindong Wu and Xingquan Zhu
• KDD05

2
Introduction

• Main challenge of mining data stream
• Data grow without limit, hard to retain long
  history
• Underlying concept may change over time
• The paper set in context of classification
  learning

3
Introduction

• Problem for now prediction data stream
• Keep recent history of raw instance
• Discard outdated concept
• No devoted to foreseeing a bigger picture

4
Introduction

• The goals involve
• (1) organizing history of raw data into history
  of compact concepts by identifying new concepts
  as well as re-appearing historical ones
• (2) learning patterns of concept transition from
  the concept history
• (3) carrying out effective and efficient
  prediction at two levels, a general level of
  predicting each oncoming concept and a specific
  level of predicting each instances class.

5
Terminology

• Data stream
• Sequence of instances
• Each instance is vector of attribute value with
  class label
• Concept
• Represented by the learning result of
  classification algorithm

6
Terminology

• Concept change
• Concept drift
• Concept shift
• Sampling change
• Change of data distribution that lead to revising
  the current model

7
Building concept history

• Key components
• Classification algorithm
• Abstract concept form raw data
• Trigger detection algorithm
• Find instance, across underlying concept changed
  and prediction model should modified
• Conceptual equivalence measure
• Check whether a concept is historical or new
• Stable learning size
• Specifies of instances which the learned
  concept can deemed stable

8
(No Transcript)
9
Building prediction model

• window size is 10
• stable learning size is 30
• trigger error threshold is 55
• represents an instance where a stable
  trigger is detected
• represents an instance where a temporary
  trigger is detected
• represents a correctly classified instance
• represents a wrongly classified instance.

10
(No Transcript)
11
(No Transcript)
12
Proactive mode

• Predict what the new concept will be when concept
  change take place
• Prepare prediction strategies in advance
• Before trigger detection and independent of
  trigger window
• Advantage
• Quick response
• Stable prediction model

13
Proactive mode

• Concept history treat as Markov Chain, each
  distinct concept is a state
• Example a sequence of arriving weather concepts
  spring, summer, autumn, winter, spring, summer,
  hurricane, autumn, winter, spring, flood, summer,
  autumn, winter, spring, summer, autumn, winter,
  spring, summer, hurricane, autumn,...

14
Reactive mode

• Reactive mode wait until concept has changed to
  construct a prediction model on trigger instance
• It can either contemporary or historical

15
RePro

• System RePro incorporate proactive and reactive
  prediction

16
Conclusion

• Proposed novel mechanism to organized data into
  concept history
• Proposed RePro to predict for concept-changing
  data streams

17
Thank you very much