TL;DR
This paper introduces COEVOLVE, a novel model that jointly captures how information spreads through social networks and how those networks evolve in response. By treating these as intertwined temporal point processes, COEVOLVE offers a more realistic and predictive understanding of online social dynamics, outperforming existing methods in predicting both information cascades and network changes.
What Problem Does This Paper Address?
Online social networks are dynamic ecosystems where information spreads and users form connections. Traditionally, these two processes—information diffusion (e.g., retweets) and network evolution (e.g., new follower relationships)—have been studied in isolation. However, they are deeply intertwined: information spread can lead to new connections, and existing connections influence how information travels. Previous work often focused on one direction of influence or relied on empirical observations without precise temporal modeling. This paper addresses the need for a unified, generative model that captures the co-evolutionary dynamics between information diffusion and network structure, allowing for more accurate predictions and a deeper understanding of these complex interactions.
Methods
The core of COEVOLVE is a joint temporal point process model. This framework is well-suited for modeling discrete events occurring over continuous time, like tweets and link creations. The model comprises two interdependent components:
- Information Diffusion Process: Modeled as an "identity-revealing" multivariate Hawkes process. This means it not only captures how retweets are triggered by previous events but also explicitly tracks the original source of the information. This allows for modeling how the "informativeness" of a source influences its ability to gain new followers.
- Network Evolution Process: Modeled as an "information-driven" survival process. The intensity of link creation is coupled with retweeting events. Crucially, new links are targeted towards information sources, with the intensity proportional to their influence and excitation.
The model is designed for efficient simulation using a novel sampling procedure that exploits the sparsity of social networks. Parameter learning is achieved through a convex optimization framework, specifically using a Minorization-Maximization (MM) algorithm, which is parameter-free and robust to initialization.
Datasets
The paper uses a large Twitter dataset containing approximately 550,000 tweet, retweet, and link events from over 280,000 users. This dataset covers a 20-day period, with specific subsets curated for source nodes, their followers, and users who start following source nodes. For experimental purposes, they also generate synthetic data from their model.
Key Insights and Ideas
- Intertwined Dynamics are Crucial: The paper strongly demonstrates that information diffusion and network evolution are not independent but mutually influence each other. Ignoring this co-evolution leads to less accurate predictions.
- "Identity-Revealing" Diffusion: Explicitly modeling the source of information is key. This allows the model to capture how a source's perceived "informativeness" directly impacts its ability to attract new followers.
- Information-Driven Link Creation: New connections are not random but are influenced by the information flow. Users are more likely to follow sources that are actively sharing relevant content or are being retweeted by their existing network.
- Efficient Simulation and Learning: The development of a specialized simulation algorithm and a convex optimization approach for parameter learning makes the model practical for large-scale datasets.
- Flexibility in Modeling: The framework can be adapted to different types of network evolution (e.g., "information-driven" link creation based on exposure vs. explicit interaction) and can incorporate external features, making it versatile for various social network scenarios.
Key Results
- Improved Prediction Accuracy: COEVOLVE significantly outperforms state-of-the-art methods (TRF and WENG) in predicting future link creation events and user activity on both synthetic and real-world Twitter data. For instance, it achieves a success probability up to 2.5x larger than TRF and WENG for link prediction.
- Accurate Parameter Estimation: The model's parameter estimation procedure accurately recovers true parameters from historical data, with accuracy increasing with more observed events. Rank correlations for parameters like α and β reach up to 0.82.
- Realistic Network Dynamics: The model successfully generates networks exhibiting common patterns like power-law degree distributions, shrinking diameters, and varying levels of triadic closure, mimicking real-world social networks.
- Co-evolutionary Behavior: Empirical and simulated cross-covariance analyses show a strong correlation between retweet and link creation intensities, confirming the co-evolutionary dynamics captured by the model.
Limitations and Future Work
- Data Requirements: The model's performance and the ability to explore extensions (like link removal, node birth/death) depend on having comprehensive, curated datasets that capture these specific events.
- Model Complexity: While efficient, the model can become computationally intensive with very large numbers of nodes and events.
- Future Directions: The authors suggest exploring a wider range of point processes, incorporating more complex features (e.g., information overload, peer pressure), handling different time scales for diffusion and network evolution, allowing time-varying parameters, and using different triggering kernels for Hawkes processes. They also note that the current model is a showcase and many more possibilities exist within the point process framework.
Summary updated