Measuring stance dynamics in political debate using temporal graph neural networks

• Our temporal GCN (TWLGCN) model infers users’ dynamic stances with high accuracy • TWLGCN integrates multiple social relations to enhance prediction performance • TWLGCN requires minimal manual annotations for training • Model’s performance is consistent across periods, granularity levels, and user groups. Social media users’ activity reflects their preferences and emotions. Mining stances from these digital traces helps predict public opinion and voting intent, but most existing approaches treat stance as static or short-term, overlooking how opinions evolve over extended periods. We introduce TWLGCN, a temporal graph neural network (GNN) that infers and tracks user-level stance dynamically from large-scale social media data with minimal manual annotation. Using Discrete-Time Dynamic Graphs (DTDGs), built from user-hashtag interactions, and integrating auxiliary user-user networks (e.g., retweets and mentions), TWLGCN aggregates current and historical signals to predict stance for all users at each period, including those not directly engaging with the target topic. Its training objective addresses class imbalance and applies a semi-supervised contrastive loss to improve representation quality. We evaluate TWLGCN on two large Twitter datasets covering the 2020 and 2022 Chilean constitutional referendums (over 1.2 million tweets from more than 17,000 users across 20 monthly periods). At each period, TWLGCN outperforms network-based baselines in macro-F1 by more than 5% using only 0.2% labeled hashtags. Beyond predictive gains, the proposed framework delivers interpretable, multi-granular insights into stance dynamics, revealing temporal drifts and group-level correlations. These results show that TWLGCN contributes to stance inference and user modeling for dynamic social platforms and is applicable to other evolving online contexts.