GAD: General Activity Detection for Fast Clustering on Large Data

1
GAD General Activity Detection for Fast
Clustering on Large Data

• Xin Jin, Sangkyum Kim, Jiawei Han, Liangliang Cao
  and Zhijun Yin
• SDM09

2
Outline

• Introduction
• GAD
• GAD for very large clustering
• Experimental
• Conclusion
• My though

3
Introduction

• It focus on developing fast core clustering
  algorithms.
• Contribution
• Exploit activity detection for fast clustering on
  different senarios
• It can achieve very high speed than k-means

4
GAD

• Notations
• NC(i, p, j)
• pattern ps jth nearest center
• D-NC(i, p, j)
• distance from pattern p to its jth nearest center
• Dist(i, p, Cj)
• distance between pattern p and center Cj

5
Definition and Concepts

• The GAD framework function
• GAD(S, A, m, B)
• S search methods, A activity states, m the
  number of nearest center, B boundary
• Search methods
• Full search - find a patterns m nearest center
• Whole full search - perform full search for all
  the patterns
• Partial search - search from active centers
• m-search - search from a patterns previous m
  nearest centers
• 0-search - a special case of m-search
• m-boundary

6
GAD algorithm

• General algorithm
• Step 1. initialization
• Step 2. search method decision
• Step 3. update pattern ps nearest centers
  according to step 2
• Step 4. get next pattern
• Step 5. assign each pattern to its nearest center
• Step 6. go to step 2 until all the centers are
  converged

7
Exact GAD algorithm

• Search method decision
• If its previous nearest center Cprev1 is static
  at this iteration, perform partial search
• If NC(i, p, 1) become active, calculate Dist(i1,
  p, Cprev1)
• Dist(i1, p, Cprev1) lt D-NC(i, p, 1), perform
  partial search
• Dist(i1, p, Cprev1) gt D-NC(i, p, 1) or
  Dist(i1, p, Cprev1) boundary, perform partial
  search
• Dist(i1, p, Cprev1) gt D-NC(i, p, 1) and
  boundary, perform full search

8

• Update pattern ps nearest center
• If full search is decided
• Search from all centers to find the m nearest
  centers
• Update the boundary as D-NC(i1, p, m)
• If partial search is decided
• Check the patterns previous m nearest center and
  keep the static centers among them as candidates
  for the current m nearest centers
• Find current m nearest centers from the
  candidates and active centers
• If D-NC(i1, p, m) is smaller than the boundary,
  update the boundary

9
Exact GAD algorithm
10
(No Transcript)
11
Full Search
12
(No Transcript)
13
NS-AGAD (Naïve State Approximate GAD)

• Approximate GAD
• It can further accelerate the speed of E-GAD
• NS-AGAD
• If a center is static at certain iteration, it
  will continue to be static at all suture
  iterations.
• Algorithm
• If ps previous nearest center Cprev1 is static
  at this iteration, perform 0-search
• If 0-search is decided, simply copy the previous
  m nearest centers as the new m nearest centers

14
S-AGAD (Static AGAD)

• If a patterns former nearest center is static
• The area near the pattern is relatively stable
• The new nearest center will likely come from the
  patterns previous m nearest centers
• It avoid searching from other centers
• Algorithm
• If ps previous nearest center Cprev1 is static,
  perform m-search at this iteration

15
I-AGAD (Inward AGAD)

• If a patterns previous nearest center is static
  or becomes activity but moves inward to the
  patterns
• Algorithm
• If ps previous nearest center Cprev1 is static,
  perform m-search at this iteration
• If center Cprev1 is active, calculate Dist(i1,
  p, Cprev1)
• If Dist(i1, p, Cprev1) lt Dist(i, p, Cprev1),
  perform m-search
• If m-search is decided, search within previous m
  nearest centers and update the new order

16
WB-AGAD (Within-Boundary AGAD)

• It further relaxed the constraints to the
  boundary
• If a patterns previous nearest center is static
  or active at the next iteration, it is still
  within the boundary of the pattern
• The new nearest center will likely be previous m
  nearest centers

17
GAD for Very Large Clusters

• H-GAD
• Hierarchical GAD
• KD-GAD
• Kd-tree GAD
• Build two kd-tree
• Full kd-tree
• Active kd-tree

18
Experimental Evaluation
19
(No Transcript)
20
(No Transcript)
21
(No Transcript)
22
Conclusion

• Propose a General Activity Detection framework
  for fast clustering.
• It is several times faster than K-Means and the
  best speedup can be as high as 10 times.

23
My thought

• Although this paper provide new core clustering
  algorithm, but whether uses on data streaming.
• Is it the same result for different initialize
  center on GAD algorithm?