Research map / Label-efficient and robust learning / Test-time adaptation
Online Fraud Detection via Test-Time Retrieval-Based Representation Enrichment
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TRE, a lightweight plug-in that retrieves top-K neighbors from the unlabeled test set at inference (with Gaussian-weighted aggregation and adaptive-k) to counter concept drift in user-behavior-sequence fraud detection.
Problem
Concept drift in online anti-fraud systems: models trained on historical fraud behavior degrade as fraudsters' tactics evolve, and periodic retraining is costly.
Approach
Build a FAISS index over the unlabeled test-time data, retrieve top-K nearest neighbor embeddings for each query, aggregate them with Gaussian-kernel weights into the query embedding, and select the neighbor count k adaptively via a learned softmax mixture over candidate k values.
Key results
- On the fraud dataset, Transformer+TRE raises AUC from 0.7992 to 0.8233 and KS from 0.4692 to 0.4966 versus the vanilla Transformer.
- TextCNN+TRE reaches the best fraud AUC of 0.8276 and ATO AUC of 0.7079, and money laundering AUC of 0.8624 among the reported results.
- Non-parametric SimHash TRE improves Transformer fraud AUC to 0.8193 while running about twice as fast per epoch as parametric TRE (6.79 vs 15.47 GPU seconds).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TRE to TextCNN | TextCNN · Fraud Dataset | AUC: 0.8116 -> 0.8276 | Helped |
| Add TRE to TextCNN | TextCNN · ATO Dataset | AUC: 0.6720 -> 0.7079 | Helped |
| Add TRE to TextCNN | TextCNN · Money Laundering Dataset | AUC: 0.8268 -> 0.8624 | Helped |
| Use TRE instead of TTT+MAE | Transformer on Fraud Dataset · Fraud Dataset | AUC: 0.8074 -> 0.8233 | Helped |
| Use TRE instead of TTT+DAE | Transformer on Fraud Dataset · Fraud Dataset | AUC: 0.8077 -> 0.8233 | Helped |
| Add adaptive-k (ablation) | TRE with Transformer · Fraud Dataset | AUC: 0.8218 -> 0.8233 | Helped |
| Add Gaussian-kernel weighting (ablation) | TRE with Transformer · Fraud Dataset | AUC: 0.8157 -> 0.8233 | Helped |
| Add adaptive-k and Gaussian-kernel weighting (ablation) | TRE with Transformer · Fraud Dataset | AUC: 0.8154 -> 0.8233 | Helped |
| Build the retrieval index from test data instead of training data (ablation) | TRE with Transformer · Fraud Dataset | AUC: 0.8034 -> 0.8233 | Helped |
| Use SimHash-based non-parametric TRE | Transformer · Fraud Dataset | AUC: 0.7992 -> 0.8193 | Helped |
Surprising findings
- Removing the test-set retrieval index (TRE w/o T, built from training data) drops performance almost back to the Transformer baseline, indicating gains stem mainly from the test distribution.
- Current test-time training methods adapted from vision (TTT+MAE, TTT+DAE) give only limited gains on behavior-sequence fraud data.
Limitations
- Evaluated only on proprietary industrial datasets from a single mobile payment platform; no public benchmark.
- Retrieval index must be rebuilt whenever the encoder checkpoint is updated, adding overhead; parametric TRE is about 4x slower per epoch than the base Transformer.
Experimental setup
Datasets: Industrial fraud detection dataset (mobile payment platform) Industrial account takeover (ATO) dataset Industrial money laundering dataset
Benchmarks: Industrial fraud detection dataset Industrial ATO dataset Industrial money laundering dataset
Metrics: AUC KS R@P0.1
Baselines: STAN NHFM SAH-RNN SHORING Transformer Longformer Reformer RetNet TextCNN TTT+MAE TTT+DAE
Builds on: test-time training Transformer FAISS SimHash logit adjustment
Closest papers by meaning
- Pre-trained Online Contrastive Learning for Insurance Fraud Detection (AAAI 2024 · similarity 0.82)
- ScamNet: Toward Explainable Large Language Model-Based Fraudulent Shopping Website Detection (AAAI 2025 · similarity 0.82)
- On the Importance of Application-Grounded Experimental Design for Evaluating Explainable ML Methods (AAAI 2024 · similarity 0.81)
- Dynamic Neighborhood Modeling via Node-Subgraph Contrastive Learning for Graph-Based Fraud Detection (AAAI 2025 · similarity 0.81)
- Barely Supervised Learning for Graph-Based Fraud Detection (AAAI 2024 · similarity 0.81)
- Defense Against Model Stealing Based on Account-Aware Distribution Discrepancy (AAAI 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.