.png){kind=icon}
.png){kind=icon}
Return
Representative negative sampling for graph positive-unlabeled learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.131462.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• A Dynamic “Galaxy Insight" for PU Learning: We introduce a novel paradigm for negative sampling inspired not only by the static structure of galaxies, but also by their dynamics. This insight treats the positive class as a system centered around a continuously evolving “star", moving beyond traditional, static clustering-based analogies. • Dynamic Prototype-Guided Representative Negative Sampling Algorithm: We operationalize our insight through a novel sampling algorithm. Instead of calculating a fixed centroid, our method maintains a dynamically evolving positive prototype via a momentum moving average. This stable yet adaptive anchor provides a principled basis for delineating reliable negative regions and sampling representative and informative negatives. • Synergistic Representation and Sampling Framework: We propose an end-to-end framework where a self-supervised GNN first cultivates a high-quality embedding space. Within this stable space, our dynamic prototype-guided sampling mechanism is deployed to effectively identify and sample representative negatives. This synergistic design ensures that high-quality representations and sampling work in concert to train a robust classifier. • Significant Performance Improvement: On real-world datasets, StarHunter-PU not only outperforms state-of-the-art approaches but also surpasses fully labeled classifiers in most cases, validating its effectiveness and generalization capability. • Practical Application Value: This-method offers an efficient and robust binary classification solution for real-world scenarios with only positive and unlabeled data, such as social network analysis and biomedical detection.