A Real-Time Method for Identifying Abnormal Transactions Based on User Behavior Sequence Analysis

By combining time-aware location coding and cross-gated attention units, the challenge of multimodal data fusion is solved, enabling real-time identification of complex fraudulent activities and improving the risk identification capability of financial transactions.

CN122089464APending Publication Date: 2026-05-26HANGZHOU YIYATONG TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
HANGZHOU YIYATONG TECH CO LTD
Filing Date
2025-12-31
Publication Date
2026-05-26

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively identify complex fraudulent activities, particularly in their inability to capture the micro-features of user actions and process the heterogeneity of multimodal data. This results in high false negative rates for risk control models in financial transactions, and low accuracy when fraud samples are scarce.

Method used

By introducing time-aware location encoding and cross-gated attention units, combined with a contrastive anomaly loss function, we achieve deep integration of physical behavior and business logic data, capture the micro-temporal rhythm of user operations, and perform global dependency modeling in the Transformer encoder to generate multimodal fusion feature sequences for risk scoring.

Benefits of technology

It enables real-time identification of highly concealed abnormal transactions, reduces the false negative rate, improves the accuracy of identifying complex fraudulent behaviors, and can effectively detect unseen attack patterns.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122089464A_ABST
    Figure CN122089464A_ABST
Patent Text Reader

Abstract

This invention provides a real-time abnormal transaction identification method based on user behavior sequence analysis, comprising the following steps: collecting physical behavior data streams and logical business data streams of users during interaction; calculating the time difference between adjacent actions in the physical behavior sequence and the logical business sequence respectively; generating time-aware position codes, which are superimposed on the processed physical behavior sequence and logical business sequence respectively to obtain physical feature vector sequences and logical feature vector sequences; inputting them into a cross-gated attention unit to generate a cross-context vector; using fusion coefficients to perform weighted summation of the logical feature vector sequence and the cross-context vector to generate a multimodal fusion feature sequence; inputting it into a pre-set Transformer encoder for global dependency modeling, outputting a final risk score, and executing the corresponding transaction control instructions. This invention can capture the user's micro-psychological state during operation in real time and effectively identify highly concealed risks.
Need to check novelty before this filing date? Find Prior Art