Rigorous evaluation against multiple baselines on 100K timesteps with 3K annotated anomalies
| Model | Precision | Recall | F1 Score | AUC-ROC | Improvement |
|---|---|---|---|---|---|
| Isolation Forest | 72% | 68% | 70% | 78% | 21.0% |
| LSTM-AE | 79% | 82% | 81% | 86% | 10.0% |
| Transformer-AE | 84% | 85% | 84% | 90% | 7.0% |
| Our Method | 92% | 90% | 91% | 94% | BEST |
All four loss components converge smoothly, demonstrating stable training dynamics. Multi-component approach prevents mode collapse.
Radar chart reveals our method's balanced strength across all metrics. No single metric is weak — consistent excellence.
Results across 5 random seeds (mean ± std). Demonstrates reproducibility and robustness.
F1 Score
91.1% ± 0.3%
Very stable
AUC-ROC
93.8% ± 0.2%
Excellent consistency
Precision
92.3% ± 0.4%
Reliable performance
Recall
90.1% ± 0.3%
Consistent detection
Bidirectional LSTM encoder alone
Transformer encoder alone
LSTM + Transformer with fixed threshold
Full model with adaptive thresholds
Key Finding: The adaptive threshold network contributes +5.1% F1 improvement over fixed-threshold approach. LSTM + Transformer combination is necessary — neither alone reaches 90% F1.
| Model | Model Size (MB) | Latency (ms) | Throughput (samples/s) | Memory (MB) |
|---|---|---|---|---|
| Isolation Forest | 0.05 MB | 2.3 ms | 435 | 12 MB |
| LSTM-AE | 0.8 MB | 8.5 ms | 118 | 156 MB |
| Transformer-AE | 1.2 MB | 12.1 ms | 83 | 224 MB |
| Our Method | 2.1 MB | 15.3 ms | 65 | 312 MB |
Our method's latency is acceptable for near-real-time applications (15ms per sample). Trade-off between model complexity and superior accuracy is justified for production systems.