Heterogeneous Network Embeddings Using Multi-Pattern Sequence Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for analyzing relationships in heterogeneous graphs are inadequate, leading to under-sampling, limited pattern identification, and inaccurate embeddings due to the exclusion of nodes with diverse types, and are limited to analyzing single patterns at a time.
Innovation Solution
A method for generating embeddings in heterogeneous networks by selecting statistically significant patterns, creating pattern and network sequence data, and combining them to generate vectors for objects, allowing for flexible analysis and inclusion of multiple patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing techniques for analyzing heterogeneous graphs are used, then the analysis process is simpler, but the accuracy and flexibility are insufficient
Solution Approach 1:
The patent segments the heterogeneous graph analysis into multiple independent pattern types (e.g., star patterns, cycle patterns, path patterns). Each pattern type is processed separately through dedicated algorithms, then the results are aggregated to generate comprehensive embeddings. This segmentation allows complex graph analysis to be broken down into manageable components, improving both accuracy and computational efficiency.
Solution Approach 2:
The patent transforms graph structural information into multiple dimensional representations by analyzing different pattern types simultaneously. It creates embeddings that capture both local neighborhood structures and global graph topology by processing patterns of varying lengths and complexities, effectively adding dimensional richness to the representation without overwhelming complexity.
2Productivity
If techniques limiting to single pattern search are used, then the computational load is reduced, but the amount of meaningful patterns identified is limited
Solution Approach 1:
The patent implements a dynamic pattern search mechanism that adapts its behavior based on the graph structure and query requirements. The system can dynamically adjust which pattern types to search for, how deeply to traverse the graph, and which nodes to prioritize, enabling efficient identification of multiple meaningful patterns without exhaustive search of all possible subgraphs.
Solution Approach 2:
By dividing pattern identification into distinct categories (star patterns, cycle patterns, path patterns, etc.), the system can process multiple pattern types in parallel using optimized algorithms for each type. This segmentation enables the system to handle large numbers of patterns efficiently, as each pattern category can be processed independently and then combined.
3Measurement precision
If under-sampling techniques are used, then the processing time is reduced, but the accuracy of relationship representation is compromised
Solution Approach 1:
The patent applies local quality analysis by examining different regions of the heterogeneous graph with different levels of detail. It identifies and prioritizes locally significant patterns (such as dense clusters or critical bridge nodes) for intensive analysis, while applying simpler processing to less critical regions. This ensures high accuracy for important relationships without uniformly processing the entire graph at maximum detail.
Solution Approach 2:
The system performs partial analysis by focusing computational resources on the most informative patterns and nodes. It uses heuristics to identify which patterns are most likely to yield meaningful embeddings and prioritizes those for detailed processing. This partial action approach maintains high accuracy for critical relationships while reducing overall processing time compared to exhaustive analysis.
Data Source
AI summary
Provided are computer-implemented methods for generating embeddings for objects which may include receiving heterogeneous network data associated with a plurality of objects in a heterogeneous network; selecting at least one pattern of objects; determining instances of each pattern of objects based on the heterogeneous network data; generating a pattern matrix for each pattern of objects based on the instances of the pattern of objects; generating pattern sequence data associated with a portion of each pattern matrix; generating network sequence data associated with a portion of the heterogeneous network data; and combining the pattern sequence data and the network sequence data into combined sequence data. In some non-limiting embodiments or aspects, methods may include generating a vector for each object of the plurality of objects based on the combined sequence data. Systems and computer program products are also provided.