InterFuse: Interaction-Guided Adaptive Fusion of Heterogeneous Models for Payment Risk Ranking
Haoyu Jiang, Weiyu Han, Ze Zhao, Jie Sheng, Ruobing Zuo, Xiao Tan, Shuai Chen, Tianyi Zhang
Well-engineered, thoroughly ablated fusion framework with real industrial gains, but the architecture is an incremental recombination of known attention/gating components and the evidence rests partly on adapted baselines and a private dataset.
Abstract
Payment platforms use view-specific models trained on transaction attributes, behavioral histories, and relational structures to rank transactions under limited review budgets. However, existing fusion methods often focus on prediction scores or interactions among model representations, with less attention to how these interactions should guide aggregation using each model's original prediction and input coverage. To address this limitation, we propose InterFuse, a fusion framework that uses a shared fusion state from cross-model interaction to guide transaction-adaptive aggregation. InterFuse encodes each model's representation, prediction score, and input coverage as a model-specific token, using learned states for unavailable outputs. Availability-aware attention updates these tokens and produces a shared fusion state. This state then guides aggregation over available models using their updated representations, original predictions, and input coverage. We evaluate InterFuse on an industrial payment replay and three public financial benchmarks using the same upstream models across fusion methods. InterFuse achieves the highest top-ranked recall and AUC on all four datasets, improving Recall@1% by 1.11-8.09 percentage points over the strongest baselines. Ablations show the benefits of cross-model interaction and adaptive aggregation, while industrial analyses find larger interaction gains when model predictions disagree. Additional analyses examine robustness to missing outputs and the trade-off between ranking gains and fusion latency.
AI Impact Assessments
(1 model)Core Contribution
InterFuse is a post-hoc fusion network that combines frozen, independently trained view-specific risk models (tabular, sequence, graph) for payment-risk ranking. Each model's representation, logit and input-coverage vector become one token, and a learned "missing" token covers unavailable outputs. A small Transformer applies availability-aware masked attention, with a special mask rule that lets the fusion token read missing states and lets missing tokens keep self-connections. The resulting shared fusion state is then fed into a scoring MLP that produces per-transaction aggregation weights over the updated model representations. These representations are re-concatenated with the original score and coverage. The final head takes the pooled vector and the fusion state.
The novelty is architectural and compositional: tokenizing (representation, score, coverage), applying masked attention, and conditioning pooling weights on the fusion token. Each piece is known (set transformers, MMT, gating, missing-modality masking). The contribution lies in the specific combination and in the ablation-backed argument that re-injecting original scores and coverage at both stages helps. It is a useful engineering design rather than a new paradigm.
Methodological Rigor
Strengths:
Concerns:
Potential Impact
Practical value: payment and fraud platforms commonly maintain many specialized models. A lightweight, retrainable fusion head (565K parameters, about 3 ms) that tolerates missing outputs and trains on cached outputs is directly deployable. The industrial gain of +5.4 pp Recall@1% over the best single model, and +1.7 pp over the best fusion baseline, is meaningful at payment scale. The public-benchmark gains are large on IBM CCT, but the baseline behaviour looks somewhat unusual.
Scientific impact is modest. The paper is a good incremental contribution to the fraud-detection and applied-ML literature (KDD/WWW industry-track style). It is unlikely to change general multimodal fusion practice. The main reusable ideas are the missing-state communication rule and the score/coverage re-injection, and the ablations support both of them as small but consistent gains (about 0.5–1 pp for the mask).
Timeliness & Relevance
Model composition and reuse of frozen models, along with missing-modality robustness, are active topics. The motivation of "payment-risk foundation models built from existing business models" is timely for industry. The setting of heterogeneous view-specific models is realistic and under-studied relative to image-text multimodality.
Strengths
Limitations
Overall
A solid applied fusion paper with careful ablations and a practical angle, but limited conceptual novelty and an unreleased industrial component. Likely to be cited by fraud-detection and model-fusion work as a strong recent baseline, with moderate broader influence.
Generated Nov 3, 2026
Comparison History (0)
Well-engineered, thoroughly ablated fusion framework with real industrial gains, but the architecture is an incremental recombination of known attention/gating components and the evidence rests partly on adapted baselines and a private dataset.