Online Fraud Detection via Test-Time Retrieval-Based Representation Enrichment
Authors: Yiran Qiao, Ningtao Wang, Yuncong Gao, Yang Yang, Xing Fu, Weiqiang Wang, Xiang Ao
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on three large-scale real-world datasets demonstrate the superiority of TRE. By consistently incorporating information from the nearest neighbors, TRE demonstrates high adaptability and surpasses existing methods in performance. |
| Researcher Affiliation | Collaboration | 1Key Lab of Intelligent Information Processing of Chinese Academy of Sciences (CAS), Institute of Computing Technology, CAS, Beijing 100190, China 2Key Lab of AI Safety, Chinese Academy of Sciences, Beijing 100094, China 3University of Chinese Academy of Sciences, CAS, Beijing 100049, China 4Independent Researcher EMAIL, EMAIL, EMAIL, name EMAIL, EMAIL, EMAIL, EMAIL Xiang Ao is also at CASMINO Ltd., Suzhou 215000, China. |
| Pseudocode | No | The paper describes steps and processes in paragraph text and uses diagrams (e.g., Figure 2) but does not include any clearly labeled 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | No | The paper mentions using a third-party open-source library: "For implementation, we apply the offline deployment of Facebook AI Similarity Search (FAISS) (Johnson, Douze, and J egou 2019), an open-source library designed for efficient similarity search and clustering of dense embeddings." It does not provide access to the source code for the methodology described in this paper. |
| Open Datasets | No | We collected three industrial datasets from a mobile payment platform on the premise of complying with security and privacy policies, covering fraud detection, account takeover (ATO) detection, and money laundering detection tasks. |
| Dataset Splits | Yes | The training data spans from 09/01/2022 to 09/30/2022, while the testing data spans from 02/01/2023 to 02/28/2023. We also selected data from six months after the training period as the out-of-time (OOT) testing dataset. We validate our training set by extracting 20% random samples. Since the ATO and money laundry datasets are highly imbalanced, we conducted a 1:10 undersampling on the negative examples during training. |
| Hardware Specification | Yes | All results were derived using a V100-16GB GPU, with Epoch Time measured in GPU seconds. |
| Software Dependencies | No | Our TRE model is implemented using Py Torch (Paszke et al. 2019) in the Linux environment. |
| Experiment Setup | Yes | Early stopping is applied with a patience of 5 epochs. The embedding size and the batch size are set to 128 and 512, respectively. We perform the dropout at each layer as 0.1, and the total parameter count of the Retriever and the Predictor are 50.1K, as 1% of a Transformer encoder. Adam W (Loshchilov and Hutter 2018) is used as the optimizer with an initial learning rate of 1e-4. |