jks – seminar talk
DESCRIPTION
JKS – Seminar Talk. Yanjun Qi 2005.Nov.14. Outline. Mixture of Experts Feature Experts Performance Expression expert – useful or not?. Split our feature sets into four sets call them as expert. Y. Mixture of Experts. expertP. expertE. expertS. expertF. X. M. Y. - PowerPoint PPT PresentationTRANSCRIPT
1
JKS – Seminar Talk
Yanjun Qi
2005.Nov.14
2
Outline
Mixture of ExpertsFeature ExpertsPerformanceExpression expert – useful or not?
3
Mixture of Experts
Split our feature sets into four sets call them as expert
Y
expertP expertE expertS expertF
4
Mixture of Experts
Root Gate is input dependent
Weight trained by EM Each expert’s
parameter also trained by EM
X
Y
M
M
m
MpMXYpXYp1
)(),|()|(
5
Four feature experts
GROUP INDEX DATA POSITION DATASET
1. Direct PPI Group
( 4 )
100 HMS_PCI Mass
101 TAP Mass
102 Y2H
103 Synthetic Lethal
2. Functional Group
( 114 )
21- 41 GO Molecular Function
42 – 74 GO Biological Process
75 – 97 GO Component
99 Essentiality
127-151 MIPS Protein Class
152-162 MIPS Mutant Phenotype
3. Sequence Group ( 7 )
104 Gene Neighborhood / Gene Fusion / Gene Co-occur
105 Sequence Similarity
106 – 109 Homology based PPI
110 Domain-Domain Interaction
4. Expression Group ( 37 ) 1-20 Gene Expression
98 Protein Expression
111-126 Protein-DNA TF group binding
6
Performance Comparison
logistic regression Naïve bayes Random forest Support vector machine Mixture of 4 feature experts
7
Performance Comparison – AUC score
Runs AUC (value + std )
AUC R50(value + std )
AUC R300(value + std )
LRNBRFSVMME
0.8823 0.03300.9349 0.01580.9321 0.01420.9159 0.02470.9463 0.0137
0.2866 0.07070.2486 0.04720.2688 0.04820.2585 0.06380.3080 0.0780
0.4629 0.0629 0.4354 0.0720 0.4719 0.0518 0.4468 0.0721 0.4789 0.0688
8
Performance Comparison – PreCal curve
9
Expression expert – useful or not?
Due to the last expression expert did not contribute much from its single performance,
Found to be most important in RF Gini So we change this expert to the full expert
to see how performance changes or delete this expert to see the whole performance change.
10
Expression expert – useful or not?
if we change the last expression expert to the full data, the performance get a little better
if we remove the last expression expert, the performance does not affect much
Runs AUC (value + std ) AUC R50
'lrexpertlrgate.4exps''lrexpertlrgate.4exps.e2all
'lrexpertlrgate.3exps.noexpression'
0.9463 0.01370.9421 0.01800.9323 0.0214
0.3080 0.07800.3132 0.07950.3022 0.0759
11
Delete Expert in turn
From the above comparison, we then want to see how performance changes if we delete one of the other three experts
12
Delete Expert in turn
Runs AUC (value + std ) AUC R50
'lrexpertlrgate.4exps', 'lrexpertlrgate.3exps.noE.old', 'lrexpertlrgate.3exps.noP', 'lrexpertlrgate.3exps.noF', 'lrexpertlrgate.3exps.noS', 'lrexpertlrgate.3exps.noE'
0.9463 0.0137 0.9323 0.0214 0.9244 0.0222 0.8821 0.0254 0.9459 0.0128 0.9323 0.0214
0.3080 0.0780 0.3022 0.0759 0.2310 0.0566 0.2609 0.0759 0.3191 0.0719 0.3022 0.0759
13
RF based feature GINI importance measure for the experts
EXPERT Experts Included Feature Groups
GROUPED GINI IMPORATNCE
(NORMALIZED)
1234
7, 8, 9,102, 3, 4, 6, 16, 17
11, 12, 13, 141, 5, 15
0.1791 0.5567 0.0814 0.1829
14
Expression expert – useful or not?
Expression experts related features would be useful when combining with others
It is not predictive for the PPI task alone
15
Acknowledge
Judith Klein-Seetharaman.
Ziv Bar-Joseph