Dynamic Node Classification in Temporal Graph Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing temporal-based machine learning classification models face challenges in dynamic node classification due to over-smoothing, which reduces model performance and increases computational costs related to misclassification, especially in fraud detection systems.
Innovation Solution
The proposed system employs a method for dynamic node classification in temporal-based machine learning classification models by converting discrete time dynamic graphs into time-augmented spatio-temporal graphs. This involves generating adjacency matrices based on temporal walks, creating adaptive information transition matrices, and propagating node representations across information propagation layers to improve classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If deep graph neural networks are used to capture long histories and long-distance entity interactions, then the model can capture more comprehensive graph structure information, but node embeddings become less distinguishable due to over-smoothing
Solution Approach 1:
The patent applies dynamics by introducing time-dependent adaptive information transition matrices that evolve across different time periods. Instead of using a static graph neural network, the system dynamically adjusts the information transition patterns at each time step, allowing the model to capture temporal variations in graph structure while preventing over-smoothing through time-aware adaptive mechanisms.
Solution Approach 2:
The patent changes parameters by learning adaptive information transition matrices specific to each time period rather than using fixed transition patterns. This allows the model to adapt its information flow parameters dynamically, capturing long-distance interactions without losing node distinguishability by adjusting the transition parameters based on temporal context.
2Device complexity
If traditional classification models are used, then the model structure is simpler, but false negatives increase leading to higher computational costs for misclassification
Solution Approach 1:
The patent segments the classification process into multiple time-period-specific stages, with each stage using adaptive information transition matrices tailored to that period. This segmentation allows the model to maintain higher accuracy for each specific time period while keeping individual model components manageable, avoiding the need for a single overly complex monolithic model.
Solution Approach 2:
The patent adds the time dimension to the classification model by creating separate adaptive information transition matrices for different time periods. This dimensional expansion allows the model to capture temporal patterns without requiring excessive complexity in each individual time step, achieving high reliability through temporal decomposition rather than spatial complexity.
Data Source
AI summary
Described are a system, method, and computer program product for dynamic node classification in temporal-based machine learning classification models. The method includes receiving graph data of a discrete time dynamic graph including graph snapshots, and node classifications associated with all nodes in the discrete time dynamic graph. The method includes converting the discrete time dynamic graph to a time-augmented spatio-temporal graph and generating an adjacency matrix based on a temporal walk of the time-augmented spatio-temporal graph. The method includes generating an adaptive information transition matrix based on the adjacency matrix and determining feature vectors based on the nodes and the node attribute matrix of each graph snapshot. The method includes generating and propagating initial node representations across information propagation layers using the adaptive information transition matrix and classifying a node of the discrete time dynamic graph subsequent to the first time period based on final node representations.