Unsupervised Pattern Discovery via Dynamic Graph Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current graph embedding technologies, such as TGN, produce embeddings that are not invertible, making it difficult to decode actionable patterns from continuous-time dynamic graphs, especially in unsupervised settings, as the embedded input space lacks interpretability and is not amenable to describing patterns between source and destination nodes in a meaningful way.
Innovation Solution
An integrated machine learning pipeline is developed that constructs source and destination embeddings using techniques like TGN, applies clustering to separate noise and signal clusters, and uses a subsequent round of clustering to identify major destination clusters, allowing for the aggregation of source node embeddings to predict temporal divisions and compute probabilities of node associations, thereby providing a summary of learned patterns and their coherence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If graph embedding technologies like TGN are used to process continuous-time dynamic graphs, then predictive capabilities and performance metrics are improved, but interpretability and actionable semantics of embeddings deteriorate
Solution Approach 1:
The patent segments the embedding space by applying clustering algorithms to identify distinct pattern clusters. This divides the complex embedding space into interpretable segments representing different temporal patterns and node relationship types, making the embeddings actionable while preserving predictive capability
Solution Approach 2:
The patent introduces clustering algorithms as an intermediary between the graph embedding technology and the final pattern interpretation. This intermediary layer transforms the complex embeddings into clustered pattern representations that maintain predictive accuracy while providing actionable semantics
2Loss of information
If complex clustering operations are applied to decode embeddings, then pattern interpretability is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent applies preliminary clustering operations during the embedding processing pipeline to organize data into pattern clusters before final analysis. This preliminary organization reduces the complexity of subsequent interpretation steps by pre-grouping similar patterns together
Solution Approach 2:
The patent changes parameters such as cluster density thresholds and pattern coherence metrics to optimize the balance between interpretability and computational complexity. By adjusting these parameters, the system can adapt the level of clustering detail to match available computational resources
Data Source
AI summary
Discussed herein are devices, systems, and methods for unsupervised pattern discovery using continuous-time dynamic graphs. A method can include receiving, from a graph neural network (GNN), source node embeddings and destination node embeddings, clustering the destination node embeddings generated by the GNN resulting in first groups of destination node embeddings, removing, from the destination node embeddings, embeddings from a noise group of the first groups resulting in signal destination node embeddings, clustering the signal destination node embeddings resulting in second groups of destination node embeddings, and identifying a pattern in the destination node embeddings and source node embeddings based on the second groups of destination node embeddings, the source node embeddings, and the destination node embeddings.