Structured Data Reconstruction With Wavelet-GCN Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating new molecules using graph convolutional networks (GCN) suffer from low-pass characteristics that smooth graph data, leading to insufficient diversity and efficiency in molecule generation, and poor interpretability in decoding processes.
Innovation Solution
A structured data generation method involving acquiring structure and node feature representations, generating a hidden-layer feature representation, and training a decoder to reconstruct structured data, using a combination of GCN and wavelet transforms to enhance decoding efficiency and diversity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If graph convolutional network (GCN) is used for molecule structure generation, then the model can learn molecular graphs through message propagation, but the low-pass characteristics of GCN smooth the graph data causing loss of complete original signal and insufficient diversity of generated molecules
Solution Approach 1:
The patent segments the frequency spectrum into multiple bands using wavelet transform, separating the graph signal into different frequency components. This allows the model to process and reconstruct signals at multiple resolution levels, preventing information loss that occurs with single-scale GCN processing.
Solution Approach 2:
The patent transforms the graph signal from a single-dimensional representation into a multi-dimensional frequency-domain representation through wavelet transform. This dimensional transformation enables the model to capture both low-frequency structural patterns and high-frequency detailed features simultaneously.
2Reliability
If GCN is used for molecule structure generation, then the model can generate new molecules through feature representation learning, but the generation efficiency and diversity of molecules are insufficient
Solution Approach 1:
The patent performs wavelet transform preprocessing on the graph data before feeding it into the GCN, decomposing the signal into multiple frequency bands in advance. This preliminary action enables the model to work with pre-processed multi-scale features, improving both generation speed and diversity without sacrificing validity.
Solution Approach 2:
The patent changes the frequency domain parameters of the graph signal by applying wavelet transform, converting the input representation from spatial domain to frequency domain. This parameter transformation allows the model to explore a broader feature space, enhancing both efficiency and diversity of molecule generation.
3Reliability
If GCN is used for molecule structure generation, then the model can process molecular graphs, but the decoding process lacks interpretability and cannot reconstruct complete original signals
Solution Approach 1:
The patent introduces wavelet transform as an intermediary between the input graph data and the GCN processing. This intermediary transforms the data into a multi-frequency representation that preserves more information and enables better reconstruction, making the decoding process more interpretable by maintaining a clearer relationship between input and output.
Data Source
AI summary
Disclosed are a structured data generation method and apparatus that can acquire a structure feature representation and a node feature representation of sample structured data; generate a hidden-layer feature representation based on the structure feature representation and the node feature representation; input the hidden-layer feature representation into a to-be-trained decoder for structure reconstruction to obtain predicted structured data; and train the to-be-trained decoder based on the predicted structured data to obtain a specified decoder, the specified decoder being configured to perform structure reconstruction on inputted sampled data to obtain reconstructed structured data.