Projected%20clustering%20algorithms%20and%20their%20application%20on%20genomic%20data%20analysis

1
Projected clustering algorithms and their
application on genomic data analysis

• Probation Talk
• M.Phil. candidate Kevin Yip
• Supervisor Dr. D. Cheung
• 20th Dec 2002

2
Presentation Outline

• The research problem clustering high-dimensional
  datasets.
• Recent work in the community projected
  (subspace) clustering.
• My previous work HARP and HARP.1.
• My current and future work studying genomic
  data, comparing different algorithms, designing
  similarity measures, etc.

3
Curse of Dimensionality

• In a dataset of high dimensionality, some
  attributes may not help identify the
  characteristics of data points.
• If the number of such attributes is high, a point
  can be as close to its nearest neighbor as its
  farthest neighbor.

4
Curse of Dimensionality
A1
A2
R1
3
8
R2
4
7
R3
9
1
D(R1, R2)
1.4
D(R2, R3)
7.8
Difference
6.4

• Resolution feature selection

5
Projected Clustering

• What feature selection cannot do find different
  relevant attributes for different clusters.

6
Projected Clustering

• Clustering identify similar objects and form
  clusters.
• Projected clustering identify similar objects
  and the relevant attributes, and form clusters.

7
Projected Clustering

• Existing approaches
• Grid-based dimension selection
• Association rule hypergraph partitioning
• Context-specific Bayesian clustering
• Monte Carlo algorithm
• Projective clustering
• All produce projected clusters successfully in
  different ways.

8
Projected Clustering

• Some problems with the approaches
• Accuracy depends on hard-to-determine input
  parameters.
• Unable to determine the dimensionality of each
  cluster automatically.
• Can only use density as similarity measure.
• In addition, the algorithms have tested on few
  real datasets.

9
HARP and HARP.1

• HARP (A Hierarchical Algorithm with Automatic
  Relevant Attribute Selection for Projected
  Clustering)
• A hierarchical clustering framework with no
  pre-assumed similarity measure.
• Requires no hard-to-determine user inputs.
• Determines the relevant attributes of each
  cluster automatically.

10
HARP and HARP.1

• HARP.1 - an implementation of HARP using
  attribute value density to define the similarity
  measure
• Makes use of global statistics in attribute
  selection.
• Handles mutual disagreement and information loss
  of the potentially merging clusters.
• Accepts mixed types of attributes.

11
HARP and HARP.1

• Some experimental results (synthetic data)

Dataset d l HARP.1 PROCLUS Traditional ORCLUS
SynCat1 20 12 0.0 / 5.0 3.6 / 1.4 6.7 / 3.7 2.6 / 26.4 5.8 / 5.3 N/A
SynMix1 20 12 0.4 / 6.8 2.2 / 17.0 6.8 / 10.1 11.6 / 11.2 7.9 / 4.6 N/A
SynNum1 20 12 0.8 / 5.0 1.8 / 21.4 7.2 / 8.3 4.4 / 32.0 5.9 / 9.2 0.4 / 23.8 2.31 / 8.15
d dimensionality of the dataset l average
number of relevant attributes
Best score error / outlier
Average error / outlier
12
HARP and HARP.1

• Some experimental results (real data)

Dataset d l HARP.1 PROCLUS Traditional ROCK
Soybean 35 26 0.0 / 0.0 0.0 / 0.0 17.3 / 0.0 2.1 / 0.0 9.2 / 0.0 No published result
Voting 16 11 6.4 / 13.6 2.1 / 55.6 13.8 / 7.9 13.1 / 11.3 13.1 / 1.9 6.2 / 14.5
Mushroom 22 15 1.4 / 0.0 3.2 / 0.0 9.0 / 0.0 6.0 / 0.0 5.2 / 0.0 0.4 / 0.0
d dimensionality of the dataset l average
number of relevant attributes (determined by
HARP.1)
Best score error / outlier
Average error / outlier
13
Current and Future Work

• A real application of projected clustering -
  analyzing genomic data
• High dimensionality
• Noisy
• Correct partition not always available
• A lot of hidden information
• New data available with a high growth rate

14
Current and Future Work
15
Current and Future Work

• Codon usage data
• Study the relationship between the frequencies of
  different codons and the functions of the genes.

16
Current and Future Work

• Transcriptome data
• The full complement of activated genes, mRNAs, or
  transcripts in a particular tissue at a
  particular time.
• Study the expression of the genes
• In different samples
• Under different situations
• At different time

17
Current and Future Work

• Transcriptome data

18
Current and Future Work

• Transcriptome data

1 2 3 4 5 6 7
51 52 53 54 55 56 57






2000 0
1 2 3 4 5 6 7
51 52 53 54 55 56 57






1 2 3 4 5 6 7
51 52 53 54 55 56 57






1 2 3 4 5 6 7
51 52 53 54 55 56 57






1 2 3 4 5 6 7
51 52 53 54 55 56 57






19
Current and Future Work

• A sample clustering result

Operon 12_1 trs5_7 0 yefJ 4 wbbK 4 wbbJ 4 wbbI 4
wbbH 4 glf 4 rfbX 4 rfbC 4 rfbA 4 rfbD 4 rfbB 4
Operon 12_2 hyfA 0 hyfB 0 hyfC 0 hyfD 0 hyfE 6 hy
fF 0 hyfG 0 hyfH 0 hyfI 0 b2490 9 hyfR 0 focB 0
Operon 12_3 rpmJ 6 prlA 0 rplO 2 rpmD 2 rpsE 2 rp
lR 2 rplF 2 rpsH 2 rpsN 2 rplE 2 rplX 2 rplN 2
20
Current and Future Work

• Goal knowledge ?clustering method
• Subspace v.s. non-subspace
• Similarity density, correlation, pattern, etc.
• Usability sensitivity to input parameters
• Algorithm type hierarchical, partitional,
  graph-based, model-based, etc.
• Preprocessing logarithm, mean-centered, PCA,
  FCA, ICA, etc.

21
Current and Future Work

• Other subtasks
• Internal validation for projected clustering
• Designing classifiers based on current results
• Building an integrated system for genomic data
  analysis

22
Conclusion

• Research space both theoretical and practical.
• Algorithms
• Projected clustering algorithms have much to
  improve.
• Many open problems in high-dimensional data
  analysis.
• Genomic data
• A lot to discover. CS people really help.

23
References

• References for projected clustering can be found
  in the slides of the presentations HARP A
  Hierarchical Approach with Automatic Relevant
  Attribute Selection for Projected Clustering and
  The Subspace Clustering Problem.
• Reference web sites for the BioInformatics
  materials covered in this presentation
• Human Genome Project Information
• HKU-Pasteur Research Centre
• Companion Web Site, Concepts of Genetics (6th Ed.)

24
Source of Figures

• Projected clusters (p.5) C. M. Procopiuc, M.
  Jones, P. K. Agarwal, and T. M. Murali. A Monte
  Carlo Algorithm for Fast Projective Clustering.
  In ACM SIGMOD International Conference on
  Management of Data, 2002.
• Generalized animal cell, genetic information flow
  (p. 14, 15) William S. Klug and Michael R.
  Cummings. Concepts of Genetics, Sixth Edition (p.
  18, 350).
• DNA with features (p. 14) Human Genome Project
  Information, http//www.ornl.gov/hgmis/.

25
Source of Figures

• GeneChip Arrays for Gene Expression Analysis (p.
  17) Affymetrix, http//www.affymetrix.com/.
• An illuminated microarray (p. 17) EMBL-EMI,
  http//www.ebi.ac.uk/.

26
Thank You!