Uncovering Social Network Activity Using Joint User and Topic Interaction

This paper introduces the Mixture of Interacting Cascades (MIC), a joint user-topic interaction model based on marked multidimensional Hawkes processes that outperforms existing methods in modeling information spread and provides insightful visualizations of social network activity.

The Gist

Imagine the internet as a massive, chaotic party where thousands of people are talking, sharing memes, arguing about politics, and posting about new songs all at once.

In the past, scientists trying to understand this party had to choose between two very simple ways of looking at it:

1. The "Copycat" View: They assumed that if Person A posts something, Person B just copies them because they are friends. (This is like a game of "Telephone").
2. The "Topic" View: They assumed that if a topic is popular, everyone talks about it, regardless of who they know.

The problem is, real life is messy. People don't just copy friends; they also get influenced by what their friends are talking about. If your friend starts talking about a new movie, you might talk about it too. But if that same friend suddenly starts ranting about a political scandal, you might ignore that specific topic even though you still listen to them about movies.

This paper introduces a new tool called "MIC" (Mixture of Interacting Cascades) to understand this mess.

Here is how MIC works, explained with simple analogies:

1. The Two-Layer Cake

Imagine the social network as a two-layer cake:

• The Bottom Layer (The Topics/Cascades): This is the "content" layer. Think of it as a table with different bowls of fruit (movies, politics, music, memes).
• The Top Layer (The People/Users): This is the "people" layer. These are the guests at the party.

Old models treated these layers separately. MIC realizes that the layers are glued together. The people on top influence the fruit bowls below, but the fruit bowls also influence which people on top get excited.

2. The "Flavor Mixing" Machine

The core magic of MIC is how it handles Topic Interaction.

Imagine you are at a buffet.

• Old Model: You decide what to eat based only on who is standing next to you. If your friend grabs a burger, you grab a burger.
• MIC Model: You decide what to eat based on your friend AND the other food on the table.
  • If your friend grabs a burger, you might grab a burger too.
  • BUT, if your friend grabs a burger, and you know that burgers go great with fries, you might also grab fries (even if your friend didn't).
  • AND, if your friend grabs a spicy taco, you might decide not to grab a burger because you know spicy food makes you thirsty, and you want a soda instead.

MIC calculates these "flavor pairings." It understands that some topics reinforce each other (Burgers + Fries), while others compete (Burgers vs. Salad). It creates a map of how topics talk to each other, not just how people talk to people.

3. The "Attention Span" Clock

MIC also understands that our attention fades.

• If your friend posts something 5 minutes ago, you might see it.
• If they posted it 5 days ago, you probably forgot.

MIC uses a "fading clock" (mathematically called a Hawkes Process) to weigh recent events more heavily than old ones. It knows that a viral tweet from an hour ago is more likely to make you retweet than one from last week.

4. Why This Matters (The Results)

The authors tested MIC on real data, like Twitter feeds during the French election and music listening habits.

• Better Prediction: MIC was much better at predicting what would happen next compared to older models. It could guess not just who would post, but what they would post about.
• Seeing the Hidden Structure: The paper shows that MIC can draw a "map" of the party.
  • In the music dataset, it showed that certain artists (like central hubs) connect different groups of fans.
  • In the political dataset, it revealed that certain political parties were actually "enemies" (competing for the same attention) even if they were on the same side of the spectrum, while others were surprisingly close allies.

The Bottom Line

Think of MIC as a super-smart party observer.

Instead of just counting how many times people high-fived (interactions) or how many songs were played (topics), MIC watches the entire dance floor. It sees that when the DJ plays a specific song, the crowd moves in a specific way, and that movement influences the next song the crowd requests.

By understanding that people and topics are in a constant, complex dance with each other, MIC gives us a much clearer picture of how information, opinions, and trends actually spread in our digital world.

Technical Summary

Here is a detailed technical summary of the paper "Uncovering Social Network Activity Using Joint User and Topic Interaction" by Abel et al.

1. Problem Statement

The paper addresses the challenge of modeling information diffusion in Online Social Networks (OSNs). Existing models often suffer from three main limitations:

• Decoupled Dynamics: They typically treat user behavior and information cascades (topics) separately, ignoring the complex interplay where user activity drives cascade spread and cascade topics influence user behavior.
• Oversimplified Interactions: Many models assume independent cascades or only consider simple correlations, failing to capture non-trivial, asymmetric, and context-dependent interactions between multiple topics (e.g., how a political topic might reinforce or inhibit another).
• Lack of Interpretability: Recent deep learning approaches prioritize predictive accuracy but offer limited insight into the underlying mechanisms driving observed patterns.
• Heterogeneity: Macroscopic models often overlook the strong heterogeneity in user activity volumes and the resulting uneven information pathways.

The goal is to develop a model that jointly captures user-to-user, cascade-to-cascade, and cascade-to-user interactions to better explain and predict information diffusion.

2. Methodology: The Mixture of Interacting Cascades (MIC)

The authors propose the Mixture of Interacting Cascades (MIC) model, a two-layer temporal point process based on Marked Multidimensional Hawkes Processes (MMHPs).

Core Components

The model defines the intensity of events (user posts/shares) as a function of three interaction layers:

1. User-to-User Interactions ($W$): Modeled via a standard Hawkes process kernel where user $v$ influences user $u$ with weight $w_{vu}$. This captures the social influence network.
2. Cascade-to-Cascade Interactions ($\Sigma$): A context-sensitive interaction matrix where $\sigma_{sc}$ represents the influence of cascade $s$ on cascade $c$. This allows for asymmetric reinforcement or inhibition between topics.
3. Cascade-to-User Interactions (Mixing Function $f_u$): A mechanism that determines the probability of a user $u$ engaging with a specific cascade $c$ given their current activity context.
   • The model introduces a context-sensitive independent user intensity $\nu^*_u(c)$, which aggregates the baseline interest and the influence of other cascades via $\Sigma$.
   • A mixing function $f_u(c|t)$ maps these intensities to a probability distribution. The authors utilize a Boltzmann distribution (softmax) controlled by a hyperparameter $\beta$:
     $$f_u(c|t) = \frac{\exp(\beta \nu^*_u(c))}{\sum_s \exp(\beta \nu^*_u(s))}$$
   • This allows for nonlinear competition: as $\beta \to \infty$, users focus exclusively on the most popular cascade; as $\beta \to 0$, activity is distributed uniformly.

Theoretical Derivations

• Closed-Form Expressions: The authors derive closed-form expressions for the conditional intensity and the expected number of events.
• Analytical Moments: They provide differential equations for the moments of the event counts and intensity, showing that the expected intensity converges to a stationary value under specific spectral radius conditions ($\rho(W^T) < 1/\tau$).

Parameter Inference

• Maximum Likelihood Estimation (MLE): The authors derive a convex log-likelihood function for the model parameters $\Theta = (M, \Sigma, W)$.
• Alternating Optimization: Due to the complexity of joint optimization, they propose an alternating scheme:
  1. Fix user parameters ($M, W$) and optimize the cascade interaction matrix $\Sigma$.
  2. Fix $\Sigma$ and optimize user parameters ($M, W$) in parallel for each user (due to factorization of the likelihood).
• Convexity: The negative log-likelihood is proven to be a convex function of the parameters, ensuring global convergence for the optimization steps.

3. Key Contributions

1. Novel Model Architecture: Introduction of MIC, which unifies Independent Cascades (IC) and Correlated Cascades (CC) models into a single framework capable of modeling complex, asymmetric topic interactions and nonlinear user behavior.
2. Theoretical Rigor: Derivation of closed-form expressions for event intensities and counts, providing analytical insights into the stability and convergence of the diffusion process.
3. Efficient Inference: Development of a scalable, convex optimization algorithm that allows for parallel learning of user-specific parameters.
4. Bi-layered Visualization: A method to visualize social network activity as a two-layer network (users and cascades) where node positions and connections are derived directly from the learned model parameters, revealing hidden community structures and topic alignments.

4. Experimental Results

The model was evaluated on synthetic data (generated by the MIC model itself) and four real-world datasets (Twitter political data, URL sharing, and music listening logs).

• Synthetic Data: MIC consistently outperformed IC, CC, and a linear variant (linMIC) in goodness-of-fit (log-likelihood), especially when cascade interactions were complex and nonlinear ($\beta$ was high). It successfully recovered the ground-truth interaction matrix $\Sigma$.
• Real Data Performance:
  • Goodness-of-Fit: MIC achieved the best or tied-best test log-likelihood scores across all datasets (e.g., élysée2017, lastfm, url).
  • Heterogeneity: MIC demonstrated superior ability to capture the heterogeneity of user activity. While other models struggled to replicate the distribution of highly active users, MIC accurately modeled the "power-law" nature of user engagement.
  • Ablation: The results confirmed that both the cascade interaction matrix ($\Sigma$) and the nonlinear mixing function ($\beta$) are critical for modeling real-world dynamics.
• Visualization: The bi-layered visualizations of the élysée2017 dataset successfully revealed political antagonisms and alliances (e.g., the positioning of LREM vs. FN) that were not apparent in standard network visualizations, validating the model's interpretability.

5. Significance and Impact

• Bridging Theory and Data: MIC bridges the gap between theoretical modeling (Hawkes processes) and data-driven analysis, offering a model that is both mathematically tractable and empirically powerful.
• Interpretability: Unlike "black-box" deep learning models, MIC provides explicit parameters ($\Sigma, W, M$) that can be directly interpreted as social influence, topic competition, and user interest.
• Mechanism Discovery: The model allows researchers to uncover the "socio-semantic coupling" in networks—how the entanglement of information polarization and community formation drives the spread of content.
• Scalability: The proposed inference algorithm is computationally efficient, making it applicable to large-scale social network datasets.

In conclusion, the paper presents a robust framework for understanding the complex, coupled dynamics of users and topics in social networks, demonstrating that explicitly modeling the interplay between cascades and user behavior significantly improves both predictive accuracy and mechanistic insight.