Table 2:
Linear Setting: RS; randomly select tuning parameters space to search. TPR-1; true positive rate for X1. Similar for TPR-2. TS: Teacher student network. – indicates not applicable. Deep IDA + Bootstrap is when we use the bootstrap algorithm to choose the top 20 ranked variables, train Deep IDA with the top 20-ranked variables, and then use the learned model and the test data to obtain test errors.
| Method | Error (%) | TPR-1 | TPR-2 | TPR-3 |
|---|---|---|---|---|
| Setting One | ||||
| Deep IDA + Bootstrap | 24.69 | 100.00 | 95.25 | - |
| Deep IDA+ TS | 33.87 | 33.25 | 21.75 | - |
| SIDA | 20.81 | 99.50 | 93.50 | - |
| Deep CCA + TS | 33.17 | 4.25 | 3.25 | - |
| Deep CCA on selected top 20 variables | 22.95 | - | - | - |
| RKCCA | 40.1 | - | - | - |
| RKCCA on selected top 20 variables | 42.07 | - | - | - |
| SVM (Stacked views) | 31.53 | - | - | - |
| SVM on selected top 20 variables (Stacked views) | 22.03 | - | - | - |
| Setting Two | ||||
| Deep IDA + Bootstrap | 23.16 | 100.00 | 94.75 | 78.75 |
| Deep IDA + TS | 31.22 | 72.00 | 57.75 | 47.75 |
| SIDA | 19.79 | 99.75 | 99.50 | 97.25 |
| DGCCA + TS | 60.01 | 2.0 | 2.0 | 2.25 |
| DGCCA on selected top 20 variables | 57.40 | - | - | - |
| SVM (Stacked views) | 29.06 | - | - | - |
| SVM on selected top 20 variables (Stacked views) | 19.56 | - | - | - |