Abstract
This paper introduces two novel models, MTL-IC and SPM-EE, designed to maximize community influence in dynamic social media streams. By incorporating multi-topic learning and event evolution tracking, these models move beyond the static influence assumptions of traditional methods. Experimental results on Twitter data demonstrate that accounting for user expertise (Authorities and Hubs) and shifting event topics significantly improves the precision and relevance of information propagation, achieving similarity scores up to 0.812 compared to 0.177 for classic models.
Introduction
The rapid proliferation of mobile internet and online social networks has transformed how information spreads. Influence Maximization (IM)—the task of identifying a small set of "seed" nodes to trigger the largest possible cascade of information—is a cornerstone of modern digital marketing, recommendation systems, and public health communication. However, the "one-size-fits-all" approach of classic models like the Independent Cascade (IC) often fails in the complex, multi-faceted environment of social media. These models typically assume that influence is a static property of a node, ignoring the reality that a user's impact is highly dependent on the topic of conversation and the temporal context of the event.
The MTL-IC Model: Beyond Binary Influence
To move beyond binary active/inactive states, the researchers developed the Multi-Topic Learning-based Independent Cascade (MTL-IC) model. This approach recognizes that influence is not just about who you know, but what you are talking about. The model employs Latent Dirichlet Allocation (LDA) to extract topic distributions from social media data streams.
Central to MTL-IC is the Topical HITS algorithm, which classifies users into two distinct roles:
- Authorities: Users who possess deep expertise and high-quality content within a specific topic.
- Hubs: Users who excel at discovering and disseminating information from authorities to the wider network.
By calculating topic-specific authority and hub scores, MTL-IC identifies seeds that are contextually relevant. Furthermore, the model incorporates geographic location data, acknowledging that physical proximity remains a powerful driver of social interaction and information trust, even in a digital-first world.
SPM-EE: Tracking the Evolution of Events
Social media events are rarely static; they grow, merge, and pivot. A "seed" influencer chosen at the start of a news cycle may become irrelevant as the discussion shifts. The Similarity Priority Mechanism-based Event Evolution (SPM-EE) model addresses this by treating events as dynamic clusters of posts.
Using a Similarity Priority Mechanism, the model tracks the evolution of these clusters over time. It identifies "event evolution relationships"—the paths through which one topic leads to another (e.g., how an initial event E2 evolves into E3 and eventually E5). This allows the system to dynamically update the set of seed users, ensuring that the information remains in the hands of those most aligned with the current state of the conversation. This "human-driven" tracking prevents the decay in propagation efficiency that plagues static models.
Experimental Evaluation and the SIS Metric
The models were rigorously tested using a Twitter dataset comprising over 126,000 posts and 6,500 users collected over a nine-day period. To better capture the quality of influence, the authors introduced the Similarity Impact Sum (SIS) metric. Unlike traditional metrics that merely count the number of reached nodes, SIS weights each reached node by its topical similarity to the event.
The results were striking. In scenarios involving 5 to 30 distinct topics, the MTL-IC model (paired with the proposed ANS algorithm) consistently achieved higher similarity scores, peaking at 0.812. In contrast, the classic IC model paired with the KKT algorithm struggled to maintain relevance, with scores often falling below 0.200. While traditional algorithms might occasionally reach a larger raw number of users, the MTL-IC and SPM-EE models ensured that the audience reached was actually interested in the content, maximizing the "effective" reach of the campaign.
Conclusion
By integrating multi-topic learning and event evolution tracking, this research provides a more nuanced and effective framework for influence maximization. The shift from "reach" to "relevance" reflects the maturing landscape of social media analytics. Future work will focus on optimizing these models for mega-scale networks and developing real-time parameter estimation techniques to further enhance the responsiveness of influence strategies in the face of rapidly changing global data streams.
Summary updated