Real-Time Embedding Prediction Using Velocity Feature Differences
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI/ML models face challenges in analyzing large datasets in real-time applications, such as fraud detection, due to significant latency introduced by generating static embeddings that are not generated in real-time, leading to poor predictions and performance degradation over time.
Innovation Solution
A method and system for predicting a real-time embedding by computing a difference between static and real-time velocity features using a prediction model, allowing for quick and reliable real-time embedding predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If static embeddings are computed for predefined time intervals, then computational resources are reduced and processing speed is improved, but prediction accuracy deteriorates and model performance degrades over time
Solution Approach 1:
The patent implements dynamic embedding updates by computing velocity features that capture temporal changes in transaction patterns. Instead of using static embeddings for predefined intervals, the system continuously updates embeddings based on real-time velocity features, allowing the model to adapt to changing fraud patterns while maintaining processing efficiency through selective updates rather than full recomputation.
2Reliability
If real-time embeddings are generated by analyzing historical transaction data for each cardholder, then prediction accuracy is improved, but latency increases significantly
Solution Approach 1:
The patent pre-computes velocity features and stores them in a velocity store for rapid access during real-time transactions. By preparing embedding components in advance and maintaining them in an optimized storage structure, the system eliminates the need for time-consuming historical data analysis during transaction processing, achieving both accuracy and low latency.
Solution Approach 2:
The patent segments the embedding generation process into separate components: static embeddings, velocity features, and temporal features. This segmentation allows each component to be computed and stored independently, with velocity features pre-computed and stored for quick retrieval. During real-time processing, only the integration of these pre-computed components is required, dramatically reducing latency while maintaining accuracy.
3Productivity
If static embeddings are used for fraud detection transactions, then processing efficiency is improved, but the embeddings become unreliable due to dynamic nature of transactions
Solution Approach 1:
The patent implements a feedback mechanism where velocity features capture temporal changes in transaction patterns and feed back into the embedding update process. The system monitors changes in transaction velocity and uses this feedback to dynamically adjust embeddings, ensuring they remain reliable and current while maintaining processing efficiency through targeted updates based on actual changes in behavior patterns.
Data Source
AI summary
Methods and systems for predicting a real-time embedding for a real-time transaction between a cardholder and a merchant are disclosed. The method performed by a server system includes receiving an embedding generation request for the real-time transaction. The method further includes accessing a static embedding, a set of velocity features, and a set of real-time velocity features from a database associated with the server system. The method further includes computing a difference between the set of real-time velocity features and the set of velocity features. The method further includes generating, by a prediction model associated with the server system, a real-time embedding prediction for the real-time transaction based, at least in part, on the static embedding, the set of velocity features, and the computed difference.