Heterogeneous Network Embedding via Segmentation and Attention Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing representation learning methods for homogeneous information networks are not directly applicable to heterogeneous information networks, which require the extraction and fusion of multidimensional information while capturing embedding uncertainty caused by various attributes.
Innovation Solution
A computer-implemented method that involves obtaining a bipartite network and multiple multi-view homogeneous networks, performing embedding learning processes for each network, and predicting association relationships between nodes of different object types by combining and fusing node embeddings using an attention mechanism.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If homogeneous information networks are used for representation learning, then the network structure can be simplified and computation can be easier, but information completeness is lost and heterogeneity of objects and relationships cannot be distinguished
Solution Approach 1:
The patent segments the heterogeneous information network into multiple homogeneous sub-networks based on different object types or relationship types. Each sub-network maintains the simplicity of homogeneous networks while collectively representing the full heterogeneity of the original network, thus resolving the contradiction between computational ease and information completeness.
Solution Approach 2:
The patent merges multiple homogeneous sub-networks through a fusion mechanism that integrates their respective node embeddings. This combining process restores the heterogeneity information that was separated into different sub-networks, achieving both computational simplicity in individual networks and information completeness in the fused result.
2Device complexity
If existing representation learning methods are applied to heterogeneous information networks, then the method can be simpler to implement, but the accuracy of association prediction deteriorates due to inability to handle heterogeneity
Solution Approach 1:
The patent divides the heterogeneous network into multiple homogeneous sub-networks, allowing the use of simple homogeneous network representation learning methods for each sub-network. This segmentation enables simple method implementation while maintaining prediction accuracy through the systematic handling of heterogeneity via multiple views.
Solution Approach 2:
The patent introduces a multi-view dimension by creating multiple homogeneous sub-networks from different perspectives (different object types or relationship types). This dimensional expansion allows simple methods to achieve high accuracy by capturing heterogeneity information across multiple dimensions rather than trying to handle it in a single complex network.
3Loss of information
If multiple multi-view homogeneous networks are created and combined, then the representation learning can capture heterogeneous information, but the computational cost and process complexity increase
Solution Approach 1:
The patent segments the complex heterogeneous network into multiple manageable homogeneous sub-networks. This segmentation reduces the complexity of individual networks while collectively preserving heterogeneity information, making the overall process more manageable despite the multi-view approach.
Solution Approach 2:
The patent employs a universal fusion mechanism that can integrate multiple homogeneous sub-networks with different object types or relationship types. This multi-functional fusion approach handles various heterogeneity scenarios through a single unified process, reducing overall process complexity while preserving heterogeneous information.
Data Source
AI summary
A computer-implemented method is provided. The computer-implemented method includes obtaining a bipartite network; a first multi-view homogeneous network; and a second multi-view homogeneous network; performing an embedding learning process for the bipartite network; performing an embedding learning process for the first multi-view homogeneous network; performing an embedding learning process for the second multi-view homogeneous network; and predicting association relationships between nodes of a first object type and nodes of a second object type.