All experiments in this study were conducted on a laptop computer with Intel Core(TM) i5-6300HQ CPU @ 2.30GHz x 4, 16GB of DDR3 RAM, and NVIDIA GeForce GTX 960M 4GB DDR5 GPU. The hyper-parameters used for both the proposed and the conventional models were assigned by hand, and through hyper-parameter optimization/tuning.
| Hyperparameters | GRU+SVM | GRU+Softmax |
|---|---|---|
| BATCH_SIZE | 256 | 256 |
| CELL_SIZE | 256 | 256 |
| DROPOUT_RATE | 0.85 | 0.8 |
| EPOCHS | 5 | 5 |
| LEARNING RATE | 1e-5 | 1e-6 |
| SVM_C | 0.5 | n/a |
| Variable | Training results | Testing results |
|---|---|---|
| False positive | 3017548 | 32255 |
| False negative | 487175 | 582105 |
| True positive | 5031465 | 731365 |
| True negative | 955012 | 757315 |
| Accuracy | 63.073973786244097% | 70.78705112598904% |
| Variable | Training results | Testing results |
|---|---|---|
| False positive | 889327 | 192635 |
| False negative | 862419 | 140535 |
| True positive | 4656221 | 1172935 |
| True negative | 3083233 | 596935 |
| Accuracy | 81.54347184760621% | 84.15769552647596% |
The graph below summarizes the training accuracy of GRU-SVM and GRU-Softmax:
The proposed GRU-SVM model was able to finish its training in 16 minutes and 43 seconds. On the other hand, the conventional GRU-Softmax model was able to finish its training in 17 minutes and 11 seconds. Both trainings consist of 37,075 steps (1,898,240 * 5 mod 256).
The graph below summarizes the testing accuracy of GRU-SVM and GRU-Softmax:
The proposed GRU-SVM model was able to finish its training in 1 minute and 22 seconds. On the other hand, the conventional GRU-Softmax model was able to finish its training in 1 minute and 40 seconds. Both testings consist of 8,215 steps (2103040 * 5 mod 256).