Graph Knowledge Distillation for Real-Time Risk Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing Graph Neural Networks (GNN) student models struggle to accurately predict online risky events due to a knowledge gap between their temporal view of ongoing events and the omni-view knowledge graph of past events held by teacher models, limiting their effectiveness in fraud prediction and risk detection.
Innovation Solution
Implementing graph knowledge distillation to train GNN student models using embedding vectors from both omni-view and temporal-view knowledge graphs, applying loss functions to bridge this gap and enhance their predictive capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If GNN student models use only temporal view of ongoing events, then they can operate in real-time for online prediction, but they cannot accurately predict risky events due to lack of knowledge from past events
Solution Approach 1:
The patent creates embedding vectors that copy the knowledge representation from the omni-view knowledge graph into a form that can be integrated with temporal-view data. The student model learns to process both omni-view embedding vectors (capturing historical patterns) and temporal-view embedding vectors (capturing ongoing events), effectively copying relevant historical knowledge into the prediction framework without requiring direct access to the full omni-view graph during real-time operation.
2Productivity
If GNN student models are trained without omni-view knowledge graph, then training is simpler and faster, but prediction accuracy for fraud detection deteriorates
Solution Approach 1:
The patent performs preliminary action by pre-computing embedding vectors from the omni-view knowledge graph that capture historical fraud patterns and risk indicators. These embedding vectors are prepared in advance and then used during the training of student models, allowing the models to learn from historical knowledge without requiring complex real-time access to the full omni-view graph, thus balancing training efficiency with prediction accuracy.
3Reliability
If GNN student models access full omni-view knowledge graph, then they can learn complete historical patterns, but real-time prediction performance deteriorates due to computational complexity
Solution Approach 1:
The patent extracts essential knowledge from the omni-view knowledge graph by computing embedding vectors that capture the most important historical patterns and relationships. Instead of requiring student models to access or process the entire omni-view graph during real-time prediction, the system extracts and stores condensed embedding representations that can be quickly integrated with temporal-view data, maintaining knowledge learning completeness while enabling fast real-time prediction.
Data Source
AI summary
Various embodiments described herein support or provide operations including identifying a machine-learning (ML) model associated with an omni-view knowledge graph; generating an embedding vector that represents the omni-view knowledge graph; identifying a ML model associated with a temporal-view knowledge graph; generating an embedding vector that represents the temporal-view knowledge graph; and training a ML model based on the generated embedding vectors.