Real-Time Payment Transaction Embeddings for Cold-Start Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for learning real-time payment (RTP) transaction embeddings fail to incorporate transaction-level information, leading to incomplete data for new customers and inefficiencies in machine learning models, and often require domain knowledge to select relevant features.
Innovation Solution
A method involving a first machine learning model with an embedding layer and hidden layer to convert remaining attributes into vectors, followed by inputting these vectors into a second model to perform a different task, such as fraud detection, without requiring domain knowledge.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If only customer, client, or account level embeddings are used, then the embedding process is simple, but the embeddings do not include transaction level information resulting in cold start for new customers
Solution Approach 1:
The patent merges customer-level embeddings with transaction-level embeddings to create comprehensive transaction representations. The hybrid model combines account embeddings (capturing customer behavior patterns) with transaction-specific embeddings (capturing individual transaction characteristics), thereby preserving both customer context and transaction details while avoiding cold start problems for new customers.
Solution Approach 2:
The embedding model is designed to serve multiple functions simultaneously: it captures customer behavior patterns across transactions, encodes individual transaction characteristics, and provides representations suitable for various downstream tasks including fraud detection, risk management, and behavioral analysis. This multi-functional design eliminates the need for separate embedding systems for different purposes.
2Measurement precision
If domain knowledge is required to select relevant features, then feature selection accuracy improves, but the process becomes more complex and time-consuming
Solution Approach 1:
The patent employs self-service mechanisms where the embedding model automatically learns and selects relevant features from raw transaction data without requiring manual domain knowledge input. The model performs self-supervised learning by identifying patterns and relationships in the data autonomously, thereby eliminating time-consuming manual feature selection while maintaining high accuracy through automated feature extraction and selection algorithms.
Data Source
AI summary
Methods, systems, and computer program products for learning continuous embedding space of real time payment (RTP) transactions are provided. A method may include receiving RTP data including a plurality of attributes, including a sender and a receiver. One attribute is selected as a target attribute. The remaining attributes are input into a first machine learning model (e.g., NLP model), comprising at least one embedding layer and one hidden layer, which is trained to predict the target attribute. After the model is trained, each of the remaining attributes are converted to a first vector using the at least one embedding layer of the machine learning model to form a first set of vectors. The first set of vectors are stored and subsequently input into a second machine learning model to perform at least one second task different than the first task.