Hebbian Graph Embeddings for Link Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing graph embedding algorithms, such as node2vec and deep convolutional embeddings, face challenges in link prediction and graph reconstruction, particularly as dimensionality increases, and do not provide significantly advantageous performance in predicting user-selected links.
Innovation Solution
A method that initializes a graph with nodes representing items in a training dataset to a multivariate normal distribution, models embeddings for each node as a non-convex Gaussian mixture of neighboring nodes' embeddings, and updates these embeddings based on transition probabilities and variance, using a Hebbian learning model with an error-free associative learning update rule.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing graph embedding algorithms (node2vec, deep convolutional embeddings) are used, then embeddings can be generated for nodes, but link prediction performance deteriorates as dimensionality increases
Solution Approach 1:
The patent changes the parameter of embedding dimensionality by implementing a hierarchical embedding structure where node representations are composed of multiple levels of embeddings. This allows the system to maintain effective link prediction performance while managing the complexity of high-dimensional data through structured composition rather than raw dimensionality increase.
Solution Approach 2:
The patent segments the embedding representation into hierarchical levels, where each node's embedding is composed of multiple sub-embeddings at different hierarchical levels. This segmentation allows the system to process complex relationships in a structured manner, improving link prediction accuracy without requiring a single high-dimensional embedding vector.
2Reliability
If existing graph embedding algorithms are used, then node representations can be learned, but performance in predicting user-selected links does not improve significantly
Solution Approach 1:
The patent implements a dynamic embedding update mechanism where embeddings are iteratively refined through multiple training iterations. The hierarchical structure allows embeddings to adapt and evolve during training, improving prediction accuracy while maintaining computational efficiency through the structured nature of the hierarchical updates.
Solution Approach 2:
The patent employs continuous iterative training where embeddings are continuously refined over multiple epochs. The hierarchical structure ensures that each iteration builds upon previous iterations, maintaining and improving prediction performance through continuous learning rather than discrete jumps in complexity.
Data Source
AI summary
Methods and systems for generating link predictions are provided. In one aspect, a method includes initializing a graph including a plurality of nodes representing selections of items in a training dataset to a multivariate normal distribution having a predetermined mean and a predetermined initial variance, the items in the training dataset comprising items in an item collection. The method includes, for each node in the graph, modeling embeddings for the node as embeddings of each neighboring node having a shared edge, with each being updated based at least in part on a transition probability and a variance. A predetermined number of iterations of updating are executed each iteration including an updated variance based on a learning rate. Based on receipt of an identification of an item from among the item collection, a plurality of predicted selections of items are identified using the embeddings for a node corresponding to the item.