Table 2 Comparative analysis of fusion strategies for multimodal biometrics.
From: PatternFusion: a hybrid model for pattern recognition in time-series data using ensemble learning
Fusion strategy | Model components | Interpretability | Computational efficiency | Adaptability | Performance (accuracy, F1-score, AUC) |
|---|---|---|---|---|---|
Early fusion | CNN + LSTM | Low | High | Static | 91.3%/90.1%/92.3% |
Late fusion | LSTM + SVM | Medium | Medium | Static | 90.8%/89.2%/91.8% |
Score-level fusion | GMM + Random Forest | Medium | Low | Static | 91.5%/89.4%/92.0% |
Attention fusion | Transformer + BiLSTM | Medium | Medium | Dynamic | 92.8%/91.5%/93.1% |
PatternFusion | BiLSTM + CNN + LightGBM | High | High | Dynamic | 96.4%/94.9%/98.1% |
