Table 4.

Ablation analysis on TITAN and TRANCOS datasets. Performance variations are reported for models with individual modules removed. The full model consistently outperforms all ablated variants in terms of Accuracy, Precision, F1 Score, and AUC.

Model	TITAN Dataset				TRANCOS Dataset
	Accuracy	Precision	F1 Score	AUC	Accuracy	Precision	F1 Score	AUC
w./o.Relational attention mechanism	86.37 ± 0.03	82.04 ± 0.02	83.65 ± 0.02	87.41 ± 0.03	86.18 ± 0.03	82.71 ± 0.02	84.26 ± 0.03	86.75 ± 0.02
w./o. Risk-oriented scene decoding	87.54 ± 0.02	84.10 ± 0.03	85.01 ± 0.02	88.16 ± 0.03	87.33 ± 0.03	83.66 ± 0.03	85.42 ± 0.02	88.07 ± 0.02
w./o. Directional local risk	85.76 ± 0.03	81.93 ± 0.03	82.78 ± 0.02	86.12 ± 0.02	84.79 ± 0.02	81.55 ± 0.02	82.31 ± 0.03	85.69 ± 0.03
Ours	89.42 ±0.02	86.15 ±0.03	87.63 ±0.02	90.33 ±0.02	88.91 ±0.03	85.92 ±0.02	87.18 ±0.02	89.87 ±0.03