INTERGEN: Ulaşım Tahmini için Performans Rehberli Özyinelemeli Ağırlık Füzyonu

Ekmekçi M.

Akıllı Ulaşım Sistemleri ve Uygulamaları Dergisi (Online), cilt.9, sa.2, ss.372-387, 2026 (TRDizin)

Özet

INTERGEN is a parameter-level neural-network fusion framework that recursively merges independently trained, architecture-compatible networks using validation-performance-based coefficients. Before each interpolation, hidden units are aligned layer-wise through Hungarian matching to address permutation symmetry, and intermediate models are re-evaluated on validation data. INTERGEN was evaluated on 2,762 Lancashire collision records for urban–rural classification and 731 daily bike-share observations for demand regression. Experiments involved 16 base neural networks, 20 paired random seeds, and 60/20/20 train/validation/test splits, with direct weight-merging and ensemble methods as baselines. For collision classification, aligned INTERGEN achieved a mean ROC-AUC of 0.8481, competing with Stacking and Greedy Model Soup. Removing permutation alignment substantially degraded classification performance. For bike-demand regression, INTERGEN achieved an RMSE of 1638.3 but underperformed Stacking and Greedy Model Soup, suggesting limitations in small-sample regression. Nevertheless, INTERGEN produces a single deployable network rather than requiring multiple models at inference. In collision classification, its footprint was 17,924 bytes versus 286,784 bytes for a 16-member Deep Ensemble, with lower inference latency. These findings indicate that alignment-guided recursive fusion can support competitive classification and efficient deployment, although predictive gains depend on task characteristics.