Architecture Search for Large-Scale Graphs via Importance Sampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for processing large-scale graph data in graph neural networks are inefficient, as they rely on sampling small subgraphs, which fail to handle the complexity and randomness of graph structures, leading to suboptimal architecture search and poor expansibility, especially for graphs with hundreds of millions of nodes.
Innovation Solution
The method involves obtaining a subgraph through local sampling, sampling neural network architectures in a pre-constructed super network using an importance sampling strategy, training these architectures via peer learning, and iteratively refining the super network to find an optimal architecture for large-scale graphs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If small subgraph sampling is used for architecture search, then computational cost is reduced, but search accuracy and reliability deteriorate due to graph structure complexity and sampling randomness
Solution Approach 1:
The patent segments the large-scale graph into multiple subgraphs through strategic sampling, then performs architecture search on each subgraph separately. This segmentation allows the system to handle computationally intractable large graphs by dividing them into manageable pieces while maintaining search reliability through careful subgraph selection and aggregation of results.
Solution Approach 2:
The patent performs preliminary actions by pre-processing the graph to identify important nodes and edges before architecture search. This preliminary analysis enables the system to focus computational resources on critical graph regions, improving search accuracy without proportionally increasing computational cost.
2Productivity
If small subgraph sampling is used for architecture search, then training speed is improved, but expansibility to large-scale graphs deteriorates
Solution Approach 1:
The patent implements a dynamic architecture search approach where the search process adapts to different graph scales. The system dynamically adjusts sampling strategies, subgraph sizes, and search depth based on the input graph characteristics, enabling it to efficiently handle graphs ranging from small to hundreds of millions of nodes without retraining.
Solution Approach 2:
The patent creates a universal architecture search framework that can handle graphs of varying scales through multi-functionality. The same core algorithm works across different graph sizes by adjusting parameters like subgraph sampling rate and search depth, making the system adaptable to both small and large-scale graphs without requiring separate specialized methods.
3Area of stationary object
If traditional graph sampling is used, then processing of large graphs is enabled, but architecture search capability is lost due to inability to process super networks
Solution Approach 1:
The patent introduces subgraphs as intermediaries between the full large-scale graph and the architecture search process. These subgraphs serve as representative samples that capture essential graph characteristics while being small enough for efficient architecture search. The search results from subgraphs are then aggregated to determine the optimal architecture for the entire graph.
Solution Approach 2:
The patent creates multiple copies of simplified graph structures (subgraphs) that replicate key properties of the original large graph. By performing architecture search on these copied representations rather than the full graph, the system maintains search capability while reducing computational complexity. The copied subgraphs preserve essential structural features needed for meaningful architecture evaluation.
Data Source
AI summary
An architecture search method and an architecture search apparatus for a large-scale graph, and a device and a storage medium are provided. The method includes: obtaining a subgraph of a large-scale graph by performing local sampling on the large-scale graph; sampling a plurality of neural network architectures in a pre-constructed super network according to a pre-customized importance sampling strategy; obtaining a plurality of trained neural network architectures by training, according to a peer learning method, the plurality of neural network architectures with the subgraph; obtaining a trained super network by iteratively executing the subgraph sampling, the architecture sampling, and the architecture training; and selecting an optimal architecture in the trained super network to process the large-scale graph to obtain a graph processing result.


