Link Polarity Prediction from Sparse and Noisy Labels via Multiscale Social Balance

TL;DR

This paper tackles robust link polarity prediction in signed networks under sparse and noisy labels by fusing multiscale social balance (MSB) with a semi-supervised training framework. The core approach, Learn2ReWeightSB, is model-agnostic and end-to-end, dynamically weighting samples labeled via microscale triads and mesoscale communities while leveraging unlabeled edges through social-balance constraints. The method includes a two-level optimization: (i) labeling unlabeled edges with MSB, and (ii) learning weights for those labels to guide SGNN training, with convergence grounded in Lipschitz-smooth loss theory. Empirical results on four real signed graphs show consistent gains in Accuracy and Macro-F1 over strong baselines, with ablations confirming the value of jointly using microscale and mesoscale SB and learning to reweight unlabeled information. The work advances robust SGNN training for signed networks and offers practical guidance for exploiting structural balance at multiple scales in real-world data.

Abstract

Signed Graph Neural Networks (SGNNs) have recently gained attention as an effective tool for several learning tasks on signed networks, i.e., graphs where edges have an associated polarity. One of these tasks is to predict the polarity of the links for which this information is missing, starting from the network structure and the other available polarities. However, when the available polarities are few and potentially noisy, such a task becomes challenging. In this work, we devise a semi-supervised learning framework that builds around the novel concept of \emph{multiscale social balance} to improve the prediction of link polarities in settings characterized by limited data quantity and quality. Our model-agnostic approach can seamlessly integrate with any SGNN architecture, dynamically reweighting the importance of each data sample while making strategic use of the structural information from unlabeled edges combined with social balance theory. Empirical validation demonstrates that our approach outperforms established baseline models, effectively addressing the limitations imposed by noisy and sparse data. This result underlines the benefits of incorporating multiscale social balance into SGNNs, opening new avenues for robust and accurate predictions in signed network analysis.

Figures (3)

• Figure 1: Depiction illustrating the use of (a) Microscale and (b) Mesoscale Social Balance to inform the link polarity prediction task. When closing a triad (a), the product of edge signs must always be equal to 1 (pos=1, neg=-1). When an edge crosses community boundaries its sign must be negative, while we assume it is positive inside the community (b).
• Figure 2: Gain in Macro-F1 for our Learn2ReWeightSB and competitor methods compared to an SDGNN trained only on $\mathcal{E}^L$, relative to the number of unlabeled edges. The performance of Learn2ReWeightSB increases gracefully as more unlabeled edges are incorporated into the framework.
• Figure 3: Learning curves across training epochs of our framework. Both $l_{\text{task}}$ and $l_{\text{SB}}$ decrease until reaching a plateau.