Graph Data Classification via Node Feature Deviation Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data acquisition methods for graph data suffer from uncertainties in node connections and inaccurate feature data, leading to low accuracy in embedded representation and classification.
Innovation Solution
A data classification method that acquires graph data, determines neighbor nodes and initial features, constructs a node similarity matrix, embeds coding features, decodes them, adjusts features based on similarity and deviation constraints, and classifies objects using an adjusted feature.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If graph data is acquired using existing data acquisition methods, then data processing can be performed, but the accuracy of embedded representation is low due to uncertain node connections and inaccurate feature data
Solution Approach 1:
The patent employs feedback mechanisms through the decoder that processes embedded coding features and provides feedback signals to adjust the embedding. The loss function calculates deviations between decoded features and initial features, using this feedback to iteratively refine the embedded representation, thereby improving accuracy while accounting for connection uncertainties
Solution Approach 2:
The patent transforms uncertain connection probabilities and inaccurate feature data into modified adjacency matrices and adjusted feature vectors. By changing parameters such as connection weights and feature values based on uncertainty measurements, the system adapts the graph structure to improve embedded representation accuracy
2Loss of information
If embedded coding is performed on initial features to obtain embedded coding features, then feature representation is achieved, but the accuracy is reduced due to information loss during encoding
Solution Approach 1:
The decoder serves as an intermediary between the embedded coding features and the original feature space. It reconstructs the initial features from the compressed encoding, and the reconstruction error provides information about the quality of compression, allowing optimization to minimize information loss while maintaining representation efficiency
Solution Approach 2:
The patent replaces traditional deterministic embedding mechanisms with probabilistic embedding that accounts for uncertainty. Instead of fixed feature transformations, it uses probability distributions and uncertainty measurements to preserve more information about the original data during the encoding process
Data Source
AI summary
A data classification method and apparatus, a device and a storage medium. A structural feature of the respective node in graph data may be determined according to a neighbor node of the respective node in the graph data through a deviation between the decoded feature obtained by decoding the embedded coding feature of the respective node in the graph data and the initial feature of the respective node, and then the embedded coding feature corresponding to the respective node is adjusted according to the decoded feature of the respective node and the structural feature of the respective node in the graph data to obtain the adjusted feature corresponding to the respective node, so that accuracy of an obtained feature of the respective node is improved, and thus accuracy of data classification may be improved.