Directed Network Embedding With Virtual Negative Edges
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing undirected network embedding methods fail to accurately capture asymmetric relationships between nodes due to the lack of consideration for edge directions, leading to inaccuracies in network representation.
Innovation Solution
A method and system that incorporate virtual negative edges to represent latent negative relationships between nodes in a directed network, using weighted regularized matrix factorization to infer negativity, and a local selection strategy to determine virtual negative edges, while modeling a signed directed network to preserve asymmetric proximity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If undirected network embedding methods are used, then the network structure can be represented, but asymmetric relationships between nodes cannot be accurately captured
Solution Approach 1:
The patent applies asymmetry by introducing directed edges with distinct source and target nodes, replacing symmetric undirected edges. The embedding method learns different representation vectors for source and target nodes, allowing the model to capture asymmetric relationships where the relationship from node A to node B differs from B to A, thus resolving the contradiction between maintaining network structure representation and capturing asymmetric relationships.
Solution Approach 2:
The patent extends the embedding space by introducing a new dimension for edge directionality. Instead of representing nodes with single vectors as in undirected methods, the patent uses separate representation vectors for source and target nodes, adding dimensional information that encodes directional relationships. This dimensional expansion enables accurate capture of asymmetric relationships while preserving overall network structure.
2Adaptability or versatility
If directed network embedding methods are used to consider edge directions, then asymmetric relationships can be captured, but accuracy in capturing relationships between nodes deteriorates
Solution Approach 1:
The patent segments the relationship representation into multiple components: source node vector, target node vector, and edge direction indicator. By dividing the embedding into these distinct segments, the model can independently learn and optimize each component's contribution to capturing asymmetric relationships, thereby improving overall accuracy while maintaining directional sensitivity.
Solution Approach 2:
The patent applies local quality by allowing different parts of the embedding space to specialize in different aspects of relationship representation. Source and target nodes have dedicated representation vectors that are optimized for their specific roles, enabling localized optimization of relationship capture accuracy while maintaining global consistency in the network embedding.
3Quantity of substance
If virtual negative edges are added to represent latent negative relationships, then network sparsity is alleviated, but network embedding complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-generating virtual negative edges based on network topology and relationship patterns before performing the main embedding task. This preliminary construction of negative edge data allows the embedding algorithm to work with a more complete graph structure without adding complexity to the core learning process, as the negative edges are prepared in advance using deterministic rules rather than requiring complex inference during training.
Data Source
AI summary
A virtual negative edge-based directed network embedding method and system are disclosed. A network embedding method performed by a network embedding system according to an embodiment may comprise the steps of: searching for a virtual negative edge representing a potentially negative relationship between nodes within a directed network; and performing network embedding for expressing the nodes within the directed network as low-dimensional vectors by using the searched virtual negative edge.