Sequence Wide Deep Learning Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current fraudulent transaction detection models based on machine learning primarily focus on static features, neglecting the longitudinal associations between transaction sequences, which significantly impacts detection precision.
Innovation Solution
A fraudulent transaction detection method and system that integrates both transverse and longitudinal dimension features using sequence wide and deep learning, involving feature mapping, self-learning models, and attention modules to generate depth feature vectors for classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If static training methods are used to process each transaction as a separate individual, then the model can process transactions independently, but the longitudinal association between transaction sequences is neglected, reducing detection precision
Solution Approach 1:
The patent transitions from processing transactions in a single horizontal dimension (individual transaction features) to incorporating a longitudinal dimension (transaction sequences over time). This is achieved by integrating both horizontal feature processing and longitudinal sequence analysis, allowing the model to capture temporal patterns and associations across multiple transactions while maintaining processing efficiency.
2Measurement precision
If sequence analysis methods focus on longitudinal relationship between sequence elements, then temporal patterns can be captured, but the learning capability in horizontal dimension for intra-sequence features is deficient, affecting detection precision
Solution Approach 1:
The patent merges two previously separate processing dimensions into a unified model: horizontal feature processing (intra-sequence features) and longitudinal sequence analysis (inter-sequence relationships). This integration ensures that both the learning capability for individual transaction features and the temporal pattern recognition are enhanced simultaneously, avoiding the deficiency of focusing on only one dimension.
3Ease of operation
If only horizontal dimension features are processed, then each transaction can be analyzed independently, but the longitudinal dimension factors that significantly impact detection results are not considered
Solution Approach 1:
The patent segments the feature processing into distinct horizontal and longitudinal components, then integrates them. The horizontal dimension processes individual transaction features independently, while the longitudinal dimension captures temporal relationships. This segmentation allows each dimension to be optimized separately before integration, maintaining operational simplicity while improving detection accuracy through comprehensive feature analysis.
Data Source
AI summary
A fraudulent transaction detection method comprises: performing feature mapping processing on each of a plurality of transaction data to generate corresponding feature vectors; converting the feature vectors of a transaction to be detected into integrated feature vectors based on a first self-learning model; respectively converting the feature vectors respectively of at least one time sequence transaction into time sequence feature vectors based on a second self-learning model; combining the integrated feature vectors and each of the time sequence feature vectors corresponding to each of the time sequence transactions to form depth feature vectors; classifying the depth feature vectors based on a third self-learning model to determine whether the transaction to be detected is a normal transaction or a fraudulent transaction.