Graph Node Labeling with Augmented Connections and Peer GNNs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing graph neural networks (GNNs) face challenges in accurately labeling nodes due to the scarcity and high cost of labeled data, sparsity of graphs, and ambiguous labels, particularly in crowd-sourced annotations, leading to inefficiencies in training and prediction accuracy.
Innovation Solution
The implementation of a system that augments graphs by adding new connections using a link prediction GNN, followed by applying two peer GNNs with a pseudo target vector and label corruption matrix to enhance accuracy, mitigating sparsity and preventing overfitting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If crowd-sourced annotation is used to label nodes in a graph, then the quantity of labeled data increases, but the accuracy of labels deteriorates due to ambiguous labels from multiple annotators
Solution Approach 1:
The patent introduces an intermediary disambiguation module that processes the ambiguous labels from multiple annotators. This module uses confidence estimates and iterative optimization to resolve conflicts between different annotators' labels, transforming the noisy crowd-sourced labels into accurate ground truth labels for training the GNN.
Solution Approach 2:
The patent implements a feedback mechanism where the system iteratively refines label accuracy by comparing predictions from the GNN with the ambiguous labels, updating confidence estimates, and re-optimizing the objective function. This feedback loop continues until convergence, progressively improving label accuracy while maintaining the benefits of crowd-sourced data quantity.
2Loss of energy
If Graph Neural Networks are trained with scarce labeled data, then training cost is reduced, but model accuracy deteriorates due to insufficient training examples
Solution Approach 1:
The patent enables the system to self-improve label accuracy through iterative optimization without requiring additional expensive human annotation. The disambiguation module uses the existing ambiguous labels and confidence estimates to automatically refine the training data quality, allowing the model to achieve high accuracy with minimal external intervention and cost.
Solution Approach 2:
The patent performs preliminary disambiguation of labels before training the GNN by maximizing an objective function that incorporates confidence estimates. This preliminary processing of the training data ensures that the model receives high-quality labels from the start, improving training efficiency and final model accuracy without requiring more labeled data.
3Adaptability or versatility
If GNNs are trained on sparse graphs with incomplete node labels, then data availability is maintained, but prediction accuracy deteriorates due to missing information
Solution Approach 1:
The patent creates a universal framework that handles multiple scenarios (complete labels, ambiguous labels, missing labels) through a single disambiguation and training process. The system can process various data quality conditions uniformly, maintaining adaptability to different graph sparsity levels while achieving accurate predictions through the objective function maximization approach.
4Adaptability or versatility
If multiple labels are assigned to a node to capture uncertainty, then label completeness improves, but label reliability deteriorates due to inability to identify the single correct label
Solution Approach 1:
The patent introduces asymmetry in how multiple labels are treated: while all candidate labels from annotators are initially considered, the disambiguation process asymmetrically weights them based on confidence estimates. The objective function maximization selectively identifies the single correct label among multiple candidates, breaking the symmetry to restore reliability while preserving the comprehensive consideration of all possibilities.
Data Source
AI summary
A system for determining more accurate labels for nodes in a graph. The system includes an electronic computing device. The electronic computing device includes an electronic processor. The electronic processor is configured to receive a graph including a plurality of nodes linked by one or more connections. The electronic processor is also configured to augment the graph by creating one or more new connections in the graph. For each node of the plurality of nodes included in the augmented graph, the electronic processor is configured to, using a first machine learning model, determine a first vector associated with the node based on the augmented graph, using a second machine learning model, determine a second vector associated with the node based on the augmented graph, and determine the more accurate label for the node based on the first vector and the second vector.