Relational Network Embedding for Anti-Money Laundering Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for identifying suspicious transactions in anti-money laundering rely on manually constructed rule models that fail to comprehensively represent subgraph patterns in multi-hop relationships, leading to incomplete information and ineffective node classification.
Innovation Solution
The method involves constructing relational networks for sample and to-be-identified data, using network embedding learning models like Node2vec to vectorize nodes, and training classification models to associate labels with these representations, enabling comprehensive data identification and classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual construction of rule models is used to identify suspicious transactions, then the system is easy to implement, but the classification accuracy is insufficient due to inability to comprehensively represent subgraph patterns in multi-hop relationships
Solution Approach 1:
The patent replaces manual mechanical construction of rule models with automated network embedding learning models (such as Node2vec) that automatically learn vector representations of nodes and subgraphs from relational network data, enabling comprehensive capture of multi-hop relationships without manual intervention
Solution Approach 2:
The patent transforms discrete relational network structures into continuous vector space representations by learning embedding vectors for nodes and subgraphs, changing the parameter representation from categorical to continuous numerical form that enables sophisticated classification
2Loss of information
If statistics-based features are used to describe subgraph patterns, then the feature construction is simple, but the information representation is incomplete as it can only describe local (one hop) relationships
Solution Approach 1:
The patent extends the representation from local one-hop neighbors to multi-hop relationships by learning embedding vectors that capture structural information from k-hop subgraphs, adding dimensional depth to the feature representation
Solution Approach 2:
The patent creates vectorized copies of subgraph structures that preserve relational information, allowing the model to process and compare complex multi-hop patterns efficiently without manually constructing detailed feature statistics
Data Source
AI summary
A computer-implemented method includes obtaining, by one or more processing devices, a training sample dataset, wherein the training sample dataset comprises at least two pieces of sample data and a sample label corresponding to each piece of sample data; constructing, by the one or more processing devices, a first relational network of the at least two pieces of sample data, and representing the at least two pieces of sample data using nodes of the first relational network; obtaining, by the one or more processing devices, a vectorized representation of each node in the first relational network based on a predetermined network embedding learning model; and training, by the one or more processing devices, a classification model based on the training sample dataset, wherein the classification model associates the sample label with the vectorized representation of each node in the first relational network.