Exploration%20of%20a%20Heuristic%20Approach%20to%20Threshold%20Learning%20in%20Adaptive%20Filtering

1
Exploration of a Heuristic Approach to Threshold
Learning in Adaptive Filtering

• Chengxiang Zhai, Peter Jansen, David A. Evans
• Claritech Corporation
• (Now Clairvoyance Corporation)

c.zhai, p.jansen, dae_at_clairvoyancecorp.com
Clairvoyance Corporation
2
Threshold Learning in Adaptive Filtering
no
doc vector
Utility Evaluation
Scoring
Thresholding
yes
profile vector
threshold
Vector Learning
Threshold Learning
Feedback Information
3
Delivery-ratio Initial Threshold Setting

• Select a threshold corresponding to a given ratio
  of acceptance/delivery
• No relevance judgments needed

Delivery-ratio 0.2 1/5
Profile vector

D1 25.5 D2 23.2 ... Dk 11.3
Scoring
Ref corpus
4
Beta-Gamma Threshold Learning

• Optimize utility function
• Exploit relevance judgments

5
Research Questions

• Learning and initial inaccuracies Can learning
  compensate for initial inaccuracies?
• Exploitation vs. exploration Does exploration
  (lowering threshold) pay off in the long run?

score
ideal adaptive
ideal fixed
actual adaptive
actual fixed
time
6
Experiment Design

• Data collection AP88 Topics 1-50
• Utility function LF1 3R - 2R-
• IDF statistics WSJ87
• Fixed profile vector (no vector learning,?0)
• Threshold updated at every accepted doc
• Varying two parameters fix others
• initial threshold level (ratio 0.002, 0.001,
  0.0005, 0.00025, 0.000125)
• learning convergence speed (gamma 0.001, 0.01,
  0.1,0.5)

7
Learning Effect 1 Correction of Inappropriate
Initial Threshold Setting
bad initial threshold without updating
bad initial threshold with updating
8
Learning Effect 2 Early Exploration Pays Off
9
Learning Effect 3 Regular Exploration Pays Off
Later
10
Tradeoff between Exploration and Exploitation
under-explore
over-explore
11
Conclusions Future Work

• Conclusions
• Learning helps correct inappropriate initial
  threshold (threshold too low)
• Exploration (deliberately lowering the threshold)
  pays off in the long run
• For a given length of document stream, the
  optimal performance is achieved at the right
  tradeoff between exploration and exploitation
• Future work
• Formalize exploitation-exploration tradeoff
  (reinforcement learning?)