Abstract
Social media graph analytics is often hindered by task-specific architectures, noisy data, and a lack of consideration for the physical dynamics of information spread. This paper introduces the Robust Propagation-aware Representation Learning (RPRL) framework, which unifies diverse social media tasks—graph classification, node classification, and link prediction—under a single architecture. RPRL utilizes a dual-encoder system and a novel kinetic-guided loss function derived from epidemiological models to enforce physical constraints on representation learning. Experimental results across six real-world datasets demonstrate that RPRL achieves state-of-the-art performance and exhibits significant zero-shot transferability, paving the way for more robust and generalizable social media foundation models.
Introduction
Social media data is inherently graph-structured, with users as nodes and interactions as edges. However, effective analysis is complicated by three primary factors: task silos, where models for rumor detection rarely generalize to bot identification; data noise, including incomplete information and spurious interactions; and missing dynamics. Most critically, existing models often ignore the underlying "kinetics" of how information moves through a network, treating propagation as a purely data-driven pattern rather than a physical process.
The RPRL Framework
The Robust Propagation-aware Representation Learning (RPRL) framework addresses these challenges by integrating domain knowledge from epidemiology with deep learning. The architecture consists of three core components:
- Data-driven Dual-Encoder: To handle varying graph scales, the model uses a Structure-aware Graph Encoder (GNN) for local dependencies and a Structure-agnostic Context Encoder (Transformer) for global interactions.
- Propagation-aware Encoder: This module simulates node activation over time using "Node Embedding Masking," which represents time steps by considering nodes within a specific hop distance from the source (ego node).
- Kinetic-guided Loss ($L_p$): This is the framework's most novel aspect. RPRL incorporates a Markov chain-based kinetic model expressed as Ordinary Differential Equations (ODEs). The model is trained to minimize the residual between its predicted state transitions and theoretical derivatives, providing physically grounded supervision that filters out noise.
Experimental Evaluation
The authors evaluated RPRL on six diverse real-world datasets covering rumor detection (DRWeibo, Weibo), social-bot detection (Twibot-22, MGTAB), and information diffusion (Christianity, Twitter).
Key results include:
- State-of-the-Art Performance: RPRL outperformed specialized baselines (e.g., BiGCN, BotRGCN) across all task types.
- Zero-shot Transferability: In rumor detection, RPRL's zero-shot performance surpassed the 5-shot performance of several baseline models, demonstrating remarkable robustness to new domains.
- Ablation Success: Removing the kinetic-guided module led to consistent performance drops. For instance, on the Twibot-22 dataset, accuracy fell from 72.61% to 68.85% without the propagation-aware components.
Discussion and Conclusion
By treating information spread as a kinetic process, RPRL establishes a "universal language" for social media graphs. The integration of physical constraints allows the model to distinguish human-like information transmission from the noise of bot spam or random interactions. While zero-shot link prediction remains a high bar due to the complexity of predicting exact user interactions in massive networks, this work points toward the development of social media foundation models pre-trained on the general kinetics of the internet.
Summary updated