Unsupervised Label Propagation Preserving Topical Context
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current graph network scoring methods, such as Bayesian models, fail to preserve the context of node labels during label propagation and cannot effectively propagate multiple non-uniform labels, relying on random walks that converge to a principal eigenvector regardless of the starting label, thus lacking topicality consideration and requiring substantial human intervention for supervised or semi-supervised techniques.
Innovation Solution
The method creates k-hop neighborhood contextual subgraphs, computes eigenvector centrality scores, and uses a mathematical decay function to propagate labels while preserving topical context, calculating attributable prestige vectors for nodes, allowing for unsupervised label prediction and reducing storage requirements by generating subgraphs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If random walk methods are used for label propagation, then the algorithm converges to a principal eigenvector, but the topical context of labels is lost and all labels converge to the same distribution
Solution Approach 1:
The patent segments the label propagation process by introducing separate propagation matrices for different label types (e.g., P_label for location labels, P_topic for topic labels). This segmentation prevents all labels from converging to the same principal eigenvector while maintaining convergence stability, as each label type follows its own propagation pattern defined by its specific matrix.
Solution Approach 2:
The patent introduces a new dimension to the propagation process by incorporating decay functions that operate differently for different label types. The decay factor λ is applied selectively based on label type and hop distance, creating an additional dimensional control mechanism that preserves topical context while maintaining convergence through the modified propagation equation.
2Measurement precision
If supervised or semi-supervised techniques are used for label propagation, then label accuracy can be improved, but substantial human intervention is required
Solution Approach 1:
The patent implements self-service by automatically determining propagation parameters and decay factors based on the intrinsic properties of the graph and label types. The system automatically selects appropriate propagation matrices and decay rates without requiring manual tuning or human expertise, achieving high label accuracy through unsupervised learning of propagation patterns.
Solution Approach 2:
The patent dynamically changes propagation parameters (such as decay factors λ and propagation matrices) based on the specific characteristics of each label type and the graph structure. These parameters are automatically adjusted during the propagation process rather than being fixed by human operators, enabling accurate label propagation without substantial human intervention.
3Quantity of substance
If k-hop neighborhood subgraphs are created for each labeled node, then storage requirements are reduced, but the complexity of managing multiple subgraphs increases
Solution Approach 1:
The patent applies universality by using a standardized subgraph structure and propagation algorithm that works across all label types and graph sizes. The same propagation framework and matrix operations are universally applied to each k-hop subgraph, reducing the complexity of managing multiple subgraphs through consistent, reusable code patterns rather than custom handling for each case.
Data Source
AI summary
One or more computer processors creating a plurality of k-hop neighborhood contextual subgraphs utilizing extracted labelled nodes from an input graph; compute an eigenvector centrality score for each node contained in each created subgraph in the plurality of k-hop neighborhood contextual subgraphs; propagate a label for each node in each subgraph in the plurality of k-hop neighborhood contextual subgraphs leveraging an aggregated mathematical decay function, preserving a topical context of the label; calculate an attributable prestige vector for each node in each subgraph in the plurality of k-hop neighborhood contextual subgraphs based on the propagated label and the computed eigenvector centrality score associated with each node in each subgraph in the plurality of k-hop neighborhood contextual subgraph; and unsupervised predict a subsequent label for one or more subsequent nodes, subgraphs, or graphs utilizing the calculated attributable prestige vectors for each node in each subgraph.