?-Clusters Capturing Subspace Correlation in a Large Data Set - PowerPoint PPT Presentation

Loading...

PPT – ?-Clusters Capturing Subspace Correlation in a Large Data Set PowerPoint presentation | free to view - id: 1674a6-ZDc1Z



Loading


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation
Title:

?-Clusters Capturing Subspace Correlation in a Large Data Set

Description:

... rate 7 and 3rd viewer rate 9, 2nd viewer probably will like this movie too ... Movie. 3. Movie2. 10/4/09. Data Mining: Concepts and Techniques. 15 ... – PowerPoint PPT presentation

Number of Views:30
Avg rating:3.0/5.0
Slides: 35
Provided by: xuehu
Category:

less

Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: ?-Clusters Capturing Subspace Correlation in a Large Data Set


1
?-Clusters Capturing Subspace Correlation in a
Large Data Set
  • Authors Yang Jiong, Wei Wang etc.(ICDE02)
  • Presenter Xuehua Shen
  • xshen_at_uiuc.edu

2
Presentation Layout
  • Overview of Clustering
  • Related Work of ?-Clusters
  • ?-Clusters Model
  • FLOC algorithm

3
Clustering
  • Clustering the process of grouping a set of
    objects into classes of similar objects
  • Similar to one another within the same cluster
  • Dissimilar to the objects in other clusters

4
Major Clustering Methods
  • Partition algorithm
  • Hierarchy algorithm
  • Density-based
  • Grid-based
  • Model-based

5
Similarity
  • Clustering the process of grouping a set of
    objects into classes of similar objects
  • But how to define similarity?

6
Similarity cont.
  • Traditional clustering model based on distance
    functions
  • Some popular ones include Minkowski distance
  • where i (xi1, xi2, , xip) and j (xj1, xj2,
    , xjp) are two p-dimensional data objects, and q
    is a positive integer
  • But strong correlations may still exist among a
    set of objects even if they are far apart from
    each other as measured by the distance function

7
Similarity cont.
  • ?-Clusters model similar when exhibiting a
    coherent pattern on a subset of dimensions
  • Can cluster objects which show shifting pattern
    or scaling pattern

8
Similarity cont.
  • Example of Coherent Pattern
  • Shifting Pattern Scaling
    Pattern

9
Subspace Clustering
  • From high dimensional clustering (problematic)
    To subspace clustering
  • Not restricted with fixed ordering of columns
    contrasted with pattern in time-series
    data
  • Challenge curse of dimensionality!

10
Subspace Clustering cont.
  • Example of subspace clustering

CH11 CH1B CH1D CH2I CH2B
CTFC3 4392 284 4108 280 228
VPS8 401 281 120 275 298
EFB1 318 280 37 277 215
SSA1 401 292 109 580 238
FUN14 2857 285 2576 271 226
SP07 228 290 48 285 224
MDM10 538 272 266 277 236
CYS3 322 288 41 278 219
CH11 CH1D CH2B

VPS8 401 120 298
EFB1 318 37 215




CYS3 322 41 219
11
Applications
  • Microarray Data Analysis in Biology
  • E-Commerce

12
Microarray Data Analysis
  • Matrix (Dense)
  • Rows Genes
  • Columns Various Samples
  • experiment conditions or
    tissues
  • Values in Matrix expression level
  • relative abundance of the mRNA of a gene
    under
  • a specific condition

13
Microarray Data Analysis cont.
  • From Scaling Pattern to Shifting Pattern
  • Red Interested Gene, Green Controlled Gene
  • Investigations show that several genes contribute
  • to a disease, which motivates researchers to
  • identify a subset of genes whose expression
    levels
  • rise and fall coherently under a subset of
    conditions

14
E-Commerce
  • Example Rating of Movies (1 lowest rate, 10
    highest rate)
  • Shifting Pattern
  • If a new movies and 1st viewer rate 7 and 3rd
    viewer rate 9, 2nd viewer probably will like this
    movie too

Movie1 Movie2 Movie 3 Movie4
Viewer1 1 2 3 6
Viewer2 2 3 4 7
Viewer3 4 5 6 9
15
Presentation Layout
  • Overview of clustering
  • Related Work of ?-Clusters
  • ?-Clusters Model
  • FLOC algorithm

16
Related Work
  • CLIQUE, ORCLUS, PROCLUS (subspace clustering)
  • Cant capture neither the shifting pattern nor
    the scaling pattern
  • Bicluster model proposed as a measure of
    coherence of genes and conditions in a submatrix
    of a DNA array

17
Bicluster
  • Model Mean squared residue score of submatrix
  • a submatrix AIJ is called a ?-biCluster if
    H(I,J)??
  • Algorithm A random algorithm to give an
    approximate answer

18
Weakness of bicluster
  • Missing Values
  • Constraints

19
Presentation Layout
  • Overview
  • Related Work
  • ?-Clusters Model
  • FLOC algorithm

20
Occupancy Threshold
  • A parameter to control the percentage of missing
    values in a submatrix
  • Ji is the specified attributes for object i in
    ?-Clusters
  • J is the number of attributes in the ?-Clusters

21
Occupancy Threshold cont.
  • Similar occupancy threshold for attribute j in
    ?-Clusters
  • Example ?0.6

1 3
4 5
3 4
1 3 3
3 4 5
3 4 4
22
Volume
  • The volume of a ?-Clusters(I,J) is the number of
    specified entries dij in (I,J)
  • Example
  • volume is 339

1 3 3
3 4 5
3 4 4
23
Base
  • Object Base
  • Attribute Base

24
Base cont.
  • ?-Clusters Base
  • For perfect ?-Clusters

25
Residue
  • Entry Residue
  • if dij is specified
  • otherwise is 0

26
Residue cont.
  • ?-Clusters Residue
  • r-residue ?-Clusters if ?-clusters residue is
    equal to or smaller than r

27
Presentation Layout
  • Overview of Clustering
  • Related Work of ?-Clusters
  • ?-Clusters Model
  • FLOC algorithm(Flexible Overlapping Clustering)

28
Flow Chart

  • Y

  • N

Generating initial clusters
Determine the best action For each row and
each column
Perform the best action sequentially
improved
29
Initial Cluster
  • Randomly Generate k initial cluster
  • Different parameters ? makes different size
    cluster

30
Choose best actions
  • For every object or attribute, there are k
    actions which can be done,
  • Choose the best action among the k candidates
    according to gain
  • Gain is the difference between original residue
    and the residue assuming the action is done on
    the cluster

31
Choose Best Actions cont.
  • Even if gain is negative sometimes
  • we do the action in order to get the global
    optimum

32
Do the actions sequentially
  • Generate the actions sequence
  • 1) the same order in all iterations
  • 2) random order sequence
  • 3) weighted random order sequence

33
Output the Best cluster
  • After some iterations, no improvement of minimum
    residue, algorithm stops and k best cluster is
    output

34
End
  • Thank you!
About PowerShow.com