Graph-Sprints Low-Latency Feature Extraction for Dynamic Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning models for dynamic graphs, especially continuous-time dynamic graphs (CTDGs), face high latency issues due to computationally costly methods like random-walk based feature extraction, which are unsuitable for real-time applications such as fraud detection in financial transactions.
Innovation Solution
The graph-sprints method computes time-aware node embeddings using single-hop operations on incoming edges, enabling low-latency feature extraction while utilizing the most up-to-date information, and approximates random-walk based features in a streaming setting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If random-walk based feature extraction is used to compute node embeddings on continuous-time dynamic graphs, then time-aware embeddings capturing multi-hop information are obtained, but computational cost increases significantly resulting in high inference latencies
Solution Approach 1:
The patent segments the complex random-walk computation into multiple simpler single-hop operations. Instead of performing complete random-walks that traverse multiple hops, the method breaks down the feature extraction into sequential single-hop updates that accumulate neighborhood information incrementally, reducing computational complexity while preserving embedding quality
Solution Approach 2:
The patent performs preliminary computation of single-hop features and stores them in advance. When node embeddings are needed, these pre-computed features are readily available, eliminating the need to perform costly multi-hop random-walks at inference time and significantly reducing latency
2Loss of information
If multiple random-walks are generated for each seed node to extract features, then comprehensive neighborhood information is captured, but computation time increases making it infeasible for high-frequency use-cases
Solution Approach 1:
The patent merges multiple random-walk computations into a single unified single-hop operation. By combining the information gathering from what would traditionally require multiple separate random-walks into one efficient pass that aggregates neighborhood features, the method maintains information completeness while dramatically improving processing throughput
Solution Approach 2:
The patent performs a single-hop operation that captures essential neighborhood information without completing full multi-hop random-walks. This partial action approach extracts the most critical features from immediate neighbors, providing sufficient information for accurate embeddings while avoiding the excessive computation of deeper walks
3Reliability
If feature extraction is performed for each incoming edge in continuous-time dynamic graphs, then up-to-date embeddings are maintained, but computational overhead becomes prohibitive for real-time applications
Solution Approach 1:
The patent segments the feature extraction process to operate independently on each incoming edge as it arrives, rather than batch-processing. This allows the system to maintain embedding freshness by immediately incorporating new edge information through efficient single-hop updates, while keeping computational overhead manageable through the simplicity of each incremental update
Data Source
AI summary
The disclosure relates to a method and system for low-latency feature extraction on continuous-time dynamic graphs (“CTDG”), in particular for training a machine-learning model. The disclosure includes feature extraction using a CTDG from a stream of transaction data, wherein each transaction is carried between a first node and a second node among a plurality of nodes, for obtaining a feature histogram for each node, each feature histogram comprising a plurality of histogram bins for one or more features, each bin comprising a count of data values falling within each bin. The method includes obtaining a first weighted average histogram from the feature histogram of the first node using a first discount factor; and obtaining a second weighted average histogram, from the feature histogram of the second node using a second discount factor. The disclosure also includes using a hashed representation of the histogram bins of the feature histogram.