Data Set For UT & Examples

a9a (bin-class)

• classes: 2
• data: 32,561 / 16,281 (testing)
• features: 123

1. LR

• loss=0.342
• precision=0.842
• auc=0.899
• trueRecall=0.562
• falseRecall=0.938

census (bin-class)

• classes: 2
• data: 22,792 / 9,769 (testing)
• features: 148
• field number: 13

1. LR

• loss=0.333
• precision=0.850
• auc=0.900
• trueRecall=0.581
• falseRecall=0.938

2. FM

• loss=0.334
• precision=0.850
• auc=0.900
• trueRecall=0.573
• falseRecall=0.940

3. Deep FM

• loss=0.323
• precision=0.853
• auc=0.904
• trueRecall=0.595
• falseRecall=0.937

4. Deep & Wide

• loss=0.323
• precision=0.853
• auc=0.904
• trueRecall=0.595
• falseRecall=0.936

5. NFM

• loss=0.322
• precision=0.853
• auc=0.906
• trueRecall=0.597
• falseRecall=0.936

6. PNN

• loss=0.322
• precision=0.853
• auc=0.905
• trueRecall=0.597
• falseRecall=0.936

w6a (bin-class)

• classes: 2
• data: 12,292 / 8,595 (testing)
• features: 300

1. LR

• loss=0.076
• precision=0.980
• auc=0.968
• trueRecall=0.676
• falseRecall=0.988

2. DNN

• loss=0.057
• precision=0.985
• auc=0.982
• trueRecall=0.826
• falseRecall=0.990

protein (multi-class/clustering)

• classes: 3
• data: 14,895 / 6,621 (testing)
• features: 357

usps (multi-class/clustering)

• classes: 10
• data: 7,291 / 2,007 (testing)
• features: 256

abalone (regression)

• data: 4,177
• features: 8

agaricus

• classes: 2
• data: 6,513 / 1,611 (testing)
• features: 127

UT Results

LinearRegTest

• dataset: abalone
• epoch: 100
• batchsize: 100
• learnrate: 0.25
• decay: 2
• reg: 0.0001
• R-Square: 0.4743

RobustRegTest

• dataset: abalone
• epoch: 100
• batchsize: 100
• learnrate: 2
• decay: 1
• reg: 0.0001
• R-Square: 0.4517

KmeansTest

• dataset: usps
• epoch: 50
• batchsize: 3000
• C: 0.3
• sample to center average dist: 5.72