AI Cleans Noisy Social Networks Better

Social networks like Bitcoin trading platforms and review sites rely on accurate relationship data, but noise from errors or manipulation often corrupts these connections. Researchers have developed RIDGE, a new AI framework that cleans both the network structure and its labels simultaneously, achieving up to 5.45% improvement in prediction accuracy across noisy datasets.

The key finding is that RIDGE effectively denoises signed graphs—networks with both positive (friendship, trust) and negative (enmity, distrust) relationships—by applying information theory to filter out irrelevant or corrupted data. Unlike previous methods that focused only on cleaning the network structure, RIDGE addresses noise in both the input features and the target labels, making it more robust to real-world data imperfections.

Methodologically, RIDGE builds on the Graph Information Bottleneck (GIB) principle, which compresses task-irrelevant information while preserving useful signals. The researchers extended GIB to handle target space denoising, creating GIB-TD. They implemented this through feature masking, which drops irrelevant node features, and substructure sampling, which selects high-confidence edges and labels. These components work together to distill clean data from noisy inputs using a reparameterization mechanism and variational approximation, enabling efficient training with supervised classification loss and information constraints.

Results from experiments on four real-world datasets—Bitcoin-OTC, Bitcoin-Alpha, Epinions, and Slashdot—show that RIDGE consistently outperforms existing methods. Under noise levels ranging from 10% to 25% (simulated by flipping edge signs), RIDGE improved Binary-F1 scores by up to 5.45% compared to baseline models. For example, on Bitcoin-OTC with 25% noise, RIDGE achieved a Binary-F1 score of 76.57, compared to 73.90 without its denoising components. The framework also demonstrated compatibility with various signed graph neural network backbones like SGCN and SNEA, enhancing their robustness without significant computational overhead.

This advancement matters because signed graphs model critical real-world systems, from e-commerce trust networks to social media interactions, where inaccuracies can lead to flawed recommendations or security risks. By improving resilience to noise, RIDGE supports more reliable applications in fraud detection, content moderation, and relationship analysis, ensuring that AI systems base decisions on cleaner, more trustworthy data.

Limitations noted in the paper include RIDGE's minor performance trade-off in noise-free conditions and its focus on GIB-based denoising without exploring alternative lightweight heuristics. Future work could combine RIDGE with other noise-tolerant techniques to further enhance performance across diverse scenarios.