Graph Conversion Device for Neural Network Input
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Graphs with indefinite attributes pose a challenge for direct input into graph neural networks due to varying attribute sets among vertices, making it difficult to satisfy the requirement of predetermined dimension vectors for attribute values.
Innovation Solution
A graph conversion device and method that converts a graph with indefinite attributes into a graph with attributes by converting attribute names and values into predetermined formats, using attribute name and value conversion units, and generating a new graph based on the original graph topology.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a graph with indefinite attributes is directly input into a graph neural network, then the graph can represent flexible and varying attribute structures, but the requirement of predetermined dimension vectors for attribute values cannot be satisfied
Solution Approach 1:
The patent transforms the indefinite attribute structure into a fixed-dimensional vector representation by changing the parameter representation method. Each attribute is converted into a vector with a predetermined dimension, allowing the graph to be input into the neural network while maintaining the ability to represent varying attribute structures through the content of the vectors rather than their dimensions.
2Ease of operation
If attribute names are converted to fixed formats, then the graph can be processed by neural networks, but the original attribute semantics may be lost or require re-learning
Solution Approach 1:
The patent performs preliminary conversion of attribute names and values into fixed-dimensional vectors before graph processing. This preliminary action encodes the semantic information into the vector representations, so that when the graph is processed by the neural network, the semantic information is already embedded and does not require re-learning.
3Manufacturing precision
If varying attribute sets are standardized to fixed dimensions, then neural network input requirements are met, but the diversity of attribute representations is reduced
Solution Approach 1:
The patent resolves this contradiction by moving the diversity information from the dimension structure to the vector content. All attributes are represented with the same fixed dimension, but the diversity of attribute representations is preserved through the varying values and distributions within those fixed-dimensional vectors, allowing both consistency for processing and diversity for representation.
Data Source
AI summary
There are provided an attribute name conversion unit which converts an attribute name of an attribute of each of a vertex and a side constituting a first graph with an attribute into a side constituting a second graph with an attribute and an attribute of the side, the attribute being composed of a first predetermined number of types of elements arranged in a first predetermined order; an attribute value conversion unit which converts an attribute value of an attribute of each of a vertex and a side constituting the first graph with an attribute into a vertex constituting the second graph with an attribute and an attribute of the vertex, the attribute being composed of a second predetermined number of types of elements arranged in a second predetermined order; and a graph-with-attribute generation unit which generates the second graph with an attribute.