Data Analysis with a Latent Variable Model

1
Data Analysis with a Latent Variable Model

• Use Nonnegative Matrix Factoctorization
• Daniel D. Lee H. S. Seung, 1999, Nature
• Experimental Setup
• Data manipulation
• For factorization, the attribute value must be
  nonnegative.
• For each gene, subtract the minimum value of the
  expression level of the gene.
• No any further normalization.
• Assume there is 2 factors, one for each class.

2

• Result
• 50 most probable features for each class
  (decreasing order)
• 5647 5997 41 5710 45 1673 7095 5728 44 6573 6463
  6208 42 43 6669 5228 929 18 1706 7096 1221 6167
  4935 4016 2185 5167 6612 6200 7101 1375 562 5709
  5715 542 282 7100 5925 6178 1393 7029 567 5934
  2798 1764 1288 6730 1929 2401 6802 5561
• 5997 45 5647 6463 7095 5728 5710 929 5709 5506 41
  1221 44 6573 18 7096 6776 42 5307 43 5715 567
  1706 2289 4016 2798 6180 6208 6772 1676 3313 1288
  911 6167 1764 5228 5937 4209 6730 5996 2726 5561
  562 2801 1549 2756 5925 1762 46 2796
• ? start index is 0.
• Clustering
• For training data 0 1 error
• For test data 12 error(s) (sample No. 54, 66)
• More to say
• In my experiment, there is no significant
  difference in performance even when normalization
  is done for each sample.
• Any other suggestions?