Low-Rank Graph Convolutional Network Training via Non-Convex Decomposition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Graph Convolutional Networks (GCNs) require a large amount of training data due to their large weight matrices, and existing low-rank GCN methods either fail to fit the initial GCN well or have slow training speeds, affecting recommendation accuracy and efficiency.
Innovation Solution
A method using a non-convex low-rank item that can be decomposed into a combination of first and second low-rank matrices is introduced, optimizing the training objective of low-rank GCNs to reduce data volume and improve parameter learning speed while ensuring accurate fitting to the initial GCN.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Graph Convolutional Network (GCN) is used for recommendation, then recommendation accuracy is improved through better utilization of relational information, but the large weight matrices require large amounts of training data
Solution Approach 1:
The patent changes the parameter of weight matrix rank from full-rank to low-rank. By decomposing the complete weight matrix into the product of two low-rank matrices (W = L1 × L2), the model reduces the number of parameters to be learned, thereby reducing the required training data volume while maintaining recommendation accuracy through the low-rank approximation of the original GCN weight matrices.
2Quantity of substance
If existing low-rank GCN methods are used to reduce data volume, then training data requirement is reduced, but the fitting to initial GCN is poor and training speed is slow
Solution Approach 1:
The patent introduces an intermediary decomposition mechanism where the complete weight matrix is represented as the product of two low-rank matrices (L1 and L2). This intermediary representation allows the model to capture the essential low-rank structure of the weight matrices while maintaining flexibility to fit the initial GCN model, thereby improving both data efficiency and fitting accuracy simultaneously.
3Quantity of substance
If existing low-rank GCN methods are used to reduce data volume, then training data requirement is reduced, but training speed becomes slow
Solution Approach 1:
The patent segments the complete weight matrix into two separate low-rank matrices (L1 and L2). This segmentation allows for more efficient computation during training, as the low-rank decomposition reduces the computational complexity of matrix operations. By training on the factorized form rather than the complete matrix, the model achieves faster convergence and improved training speed while requiring less training data.
Data Source
AI summary
The present disclosure provides a method for generating a recommendation model, a content recommendation method, and a content recommendation apparatus, and an electronic device, and relates to an artificial intelligence field and a deep learning field. The method for generating a recommendation model includes: obtaining a graph training sample set; inputting the graph training sample set into a machine learning model to train the machine learning model, in which the machine learning model includes at least one low-rank graph convolutional network, and the low-rank graph convolutional network includes a complete weight matrix composed of a first low-rank matrix and a second low-rank matrix; in which a training objective of the low-rank graph convolutional network includes a first parameter item, a second parameter item and a non-convex low-rank item; and in responding to detecting that a training end condition is met, determining the machine learning model as a recommendation model.