Website Classification via Graph Neural Network Trust Propagation

The patent performs preliminary classification of websites as trustworthy or untrustworthy before users access them. By pre-computing trust scores and propagating trustworthiness information through the web graph in advance, the system eliminates detection lag when users encounter new or previously unknown malicious sites, while maintaining reliability through rigorous classification criteria.

The patent implements local quality by allowing each user or organization to maintain their own customized trust graph and classification preferences. Instead of a single centralized allow-list, the system enables local adaptation where users can adjust trust thresholds, select which neighbor websites to consider, and customize their browsing experience while maintaining security, thus improving ease of operation without sacrificing reliability.

The patent implements self-service by automatically propagating trustworthiness scores through the web graph using graph neural networks and semi-supervised learning. The system autonomously updates classifications by leveraging the structure of hyperlink connections and machine learning models, eliminating manual intervention entirely. This maintains high accuracy through rigorous algorithmic classification while dramatically improving productivity by removing labor-intensive manual updating processes.

The patent applies segmentation by dividing the web graph into local neighborhoods around each website being classified. Instead of performing global propagation across the entire web graph which would be computationally prohibitive, the system segments the problem into manageable local regions, computing trust scores only for relevant neighbor websites within a limited depth. This maintains high detection accuracy by considering local trust relationships while reducing computational complexity to feasible levels.