Table 1:
Ablation study of VIE with different combinations of architectures on realistic synthetic datasets with 1% event rate. The oracle model has used the ground-truth model parameters to predict.
| Average AUC (standard deviation) | Average AUPRC (standard deviation) | |||||||
|---|---|---|---|---|---|---|---|---|
|
| ||||||||
| Prior | Encoder | Decoder | Prior Match | n=5k | n=20k | n=5k | n=20k | |
| VAE | Gaussian | Gaussian | MLP | True | 0.552 (0.092) | 0.674 (0.020) | 0.026 (0.010) | 0.061 (0.017) |
| VAE-GPD | mixed GPD | Gaussian | AMNN | False | 0.569 (0.062) | 0.653 (0.027) | 0.021 (0.003) | 0.035 (0.013) |
| IAF-GPD | mixed GPD | IAF | AMNN | False | 0.511 (0.021) | 0.665 (0.029) | 0.017 (0.002) | 0.025 (0.008) |
| Fenchel-GPD | mixed GPD | Implicit | AMNN | True | 0.623 (0.036) | 0.694 (0.021) | 0.037 (0.010) | 0.062 (0.026) |
| VIE | mixed GPD | IAF | AMNN | True | 0.684 (0.031) | 0.701 (0.017) | 0.050 (0.009) | 0.079 (0.025) |
|
| ||||||||
| Oracle (with 90% confidence interval) | 0.704 [0.662, 0.751] | 0.092 [0.058, 0.141] | ||||||