InterFuse: Interaction-Guided Adaptive Fusion of Heterogeneous Models for Payment Risk Ranking (arXiv (Cornell University))

Summary

• InterFuse achieves the highest top-ranked recall and AUC on all four evaluated datasets, improving Recall@1% by 1.11–8.09 percentage points over the strongest baselines.p.1
• On the Industrial dataset, InterFuse achieves a Recall@1% gain of +5.44±0.37 percentage points over the fixed B1 baseline.p.5
• Removing learned model representations h lowers Recall@1% by 5.28 points on Industrial and 4.53 points on IBM CCT.p.7

Key insights

1. InterFuse encodes representations, prediction scores, and input coverage as tokens updated through availability-aware attention to guide adaptive aggregation.p.1
2. Cross-model interaction updates individual model tokens before aggregation, improving ranking quality beyond global output summarization.p.7
3. Cross-model interaction generates higher recall gains on transactions where view-specific model predictions disagree.p.8
4. Direct access to original prediction scores and input coverage improves performance at both interaction and aggregation stages.p.7
5. InterFuse maintains positive Recall@1% gains over clean B1 ranging from +2.86 to +3.80 percentage points across single optional model removals without retraining.p.8
6. Under 50% joint removal of optional outputs, InterFuse retains a +2.91 percentage point Recall@1% advantage over clean B1 on the Industrial dataset.p.8