TL;DR
InterFuse is a novel fusion framework designed to combine multiple specialized risk models (like tabular, sequence, and graph models) into a single, highly accurate payment risk ranking. By allowing these models to "talk" to each other through a shared fusion state, InterFuse adaptively weights their contributions based on the specific context of each transaction, leading to significant performance gains on both industrial and public benchmarks.
What Problem Does This Paper Address?
In the world of online payments, fraud detection systems rely on diverse "views" of a transaction. One model might look at tabular data (transaction amount, time), another at sequential behavior (is this purchase typical for this user?), and a third at graph relationships (is this account linked to known fraudsters?).
The challenge lies in fusion: how do you combine these different assessments into one final risk score? Traditional methods often just average the scores or concatenate features. However, these approaches ignore crucial context:
- Input Coverage: A sequence model might give a low-risk score simply because it only has two days of history for a new user, not because the user is actually safe.
- Model Disagreement: If the tabular model says "safe" but the graph model says "high risk," the system needs a smart way to decide which one to trust for that specific case.
- Missing Data: Sometimes, a specific model (like a graph model) might not produce an output at all for a given transaction.
Methods
InterFuse introduces an interaction-guided adaptive fusion architecture. Instead of just looking at final scores, it treats each model's output as a "token" in a sequence, similar to how words are treated in a Large Language Model.
The process follows three main steps:
- Model Tokenization: For every available model, InterFuse packs its learned representation ($h$), its prediction score ($s$), and its "input coverage" ($c$—e.g., history length) into a single token. If a model is missing, it uses a learned "missing state" token.
- Cross-Model Interaction: These tokens pass through Transformer-style attention layers. A special fusion token acts as a central hub, gathering information from all models to create a shared fusion state. This allows a model to update its own internal representation based on what other models are seeing.
- Adaptive Aggregation: Finally, the shared fusion state "guides" the weights for each model. If the fusion state recognizes a "new user" context, it might automatically down-weight the sequence model (which has low coverage) and up-weight the tabular model.
Datasets
The authors evaluated InterFuse across four diverse financial datasets:
- Industrial (Ant International): A massive real-world payment replay dataset using five internal models (tabular, user-history, card-history, device-history, and graph).
- TransXion: A public benchmark for anti-money laundering (AML) detection.
- IBM CCT: A public dataset focused on card fraud detection.
- MBD (Multimodal Banking Dataset): A dataset for predicting next-month product purchases, testing the framework's versatility beyond fraud.
Key Insights and Ideas
- Disagreement is a Signal: One of the most striking findings is that InterFuse provides the biggest gains when individual models disagree. When models are conflicted, the "shared fusion state" is essential for resolving the ambiguity.
- Coverage Matters: By explicitly including "input coverage" (how much data a model had to work with), the system learns to be skeptical of models operating on sparse information.
- Representations > Scores: InterFuse performs better than score-only fusion because it leverages the high-dimensional features ($h$) learned by the upstream models, which often contain nuances lost in a single scalar risk score.
- Robustness to Missingness: Because it uses learned "missing states" and availability-aware masking, the system remains highly effective even when several models fail to provide an output.
Key Results
InterFuse consistently outperformed all baselines, including advanced multi-modal transformers and expert-gating models.
- Ranking Accuracy: On the primary metric, Recall@1% (the percentage of fraud caught in the top 1% of highest-risk transactions), InterFuse improved performance by 1.11 to 8.09 percentage points over the strongest baselines.
- Industrial Impact: In the real-world industrial setting, InterFuse achieved a 5.44 percentage point gain in Recall@1% over the single best-performing individual model.
- Efficiency: Despite the added complexity of attention layers, the median fusion latency was only 3.04 ms, making it viable for real-time payment processing.
Limitations and Future Work
- Latency Trade-off: While 3ms is fast, it is still slower than simple score averaging. For platforms with extreme throughput requirements, this might be a consideration.
- Frozen Upstream Models: The current framework keeps the original risk models "frozen." While this makes it easy to deploy, "end-to-end" training (updating the original models and the fusion layer together) might unlock even higher performance, albeit at a much higher computational cost.
- Foundation Models: The authors suggest that future work could involve integrating these specialized model outputs into larger "payment-risk foundation models."
Summary updated