Graph-Based Node Selection for Efficient Data Labelling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning methods rely on unsupervised learning due to the lack of labelled data, leading to inefficiencies and high resource costs for manual labelling by domain experts, and the need for continuous retraining with unlabelled data generation during the productive phase.
Innovation Solution
A computer-implemented method using a graph database representation and joint machine learning architecture, comprising a graph neural network for feature learning and a neural network for node determination, to efficiently select and label a subset of data samples, leveraging active learning strategies and reinforcement learning to optimize labelling efforts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual labelling by domain experts is used to obtain labelled data, then the quality and reliability of labelled data improve, but the time consumption and resource costs increase significantly
Solution Approach 1:
The system enables semi-automated labelling where the machine learning model performs self-labeling on unlabelled data, and domain experts only need to review and correct uncertain predictions rather than labeling all data from scratch. This self-service approach dramatically reduces manual effort while maintaining data quality.
Solution Approach 2:
A graph-based uncertainty estimation mechanism acts as an intermediary between the machine learning model and domain experts. The uncertainty estimator identifies which predictions are reliable and which need expert review, filtering and prioritizing data samples for manual labeling based on their uncertainty scores.
2Reliability
If all data samples are labelled by domain experts to ensure comprehensive training data, then the machine learning model performance improves, but the resource intensity and costs increase
Solution Approach 1:
Instead of uniformly labeling all data samples, the system applies different labeling strategies to different regions of the data space. High-uncertainty samples that critically impact model performance receive expert labeling, while high-confidence samples are automatically accepted, creating locally optimized labeling density where it matters most.
Solution Approach 2:
The system performs partial labeling by selectively labeling only the subset of data samples that are most uncertain and most valuable for model improvement. Rather than labeling all data (excessive action) or using purely unsupervised learning (insufficient action), it finds the optimal partial set that maximizes model performance per unit of labeling effort.
3Productivity
If unsupervised machine learning is used due to lack of labelled data, then the processing speed and resource efficiency improve, but the ability to incorporate domain expert feedback is lost
Solution Approach 1:
The system dynamically adjusts between unsupervised and supervised modes based on data uncertainty. For high-confidence predictions, it operates in fast unsupervised mode; for uncertain predictions, it transitions to supervised mode where domain expert feedback is incorporated. This dynamic switching maintains processing efficiency while ensuring expert coherence when needed.
Solution Approach 2:
Domain expert feedback is systematically incorporated through the uncertainty estimation pipeline. Experts review and correct uncertain predictions, and this feedback is used to retrain the model, creating a continuous improvement loop that maintains both speed and expert coherence.
Data Source
AI summary
A method including providing a graph database representation, wherein the graph database representation represents a plurality of nodes in a graph which are interconnected by respective edges, wherein each node of the plurality of the nodes represents a data sample and is assigned to at least one node feature, wherein each edge of the plurality of the edges represents a relationship between the data samples, transforming the graph database representation into a data matrix using a first machine learning algorithm suitable for graph data and an architecture as first layers of a joint machine learning architecture, determining at least one node of the plurality of nodes based on the transformed data matrix using a second machine learning algorithm and a second architecture as second layers of a joint machine learning architecture, and providing the at least one determined node. Further, a computing unit and a computer program product is provided.