🤖 AI

A Network-driven Framework for Public Event Forecasting via Dynamic Interaction Network Evolution

This paper proposes auto-ibDLM, a network-driven deep learning framework that forecasts public event evolution by modeling dynamic participant interactions through a hybrid representation learning strategy and GRU-based temporal analysis, achieving over 97% accuracy and demonstrating superior performance across 13 real-world datasets.

Original authors: Jie Wei, Yue Liu, Xiaochuan Tang, Biao Cai, Xiangtao Li, Yanmei Hu

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jie Wei, Yue Liu, Xiaochuan Tang, Biao Cai, Xiangtao Li, Yanmei Hu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the crowded digital square of modern life, public events often begin as quiet whispers before they swell into roaring crowds. Whether it is a festival, a protest, or a sudden gathering during a crisis, the speed at which people arrive and interact can determine whether an event remains a celebration or turns into a dangerous emergency. For city planners and safety officials, the ability to see this surge coming is vital. It allows them to open extra gates, redirect traffic, or call for help before a crowd becomes too large to manage. The challenge lies in the fact that these events are not static; they are living systems driven by the constant, shifting interactions between thousands of individuals. Traditional methods often look at the crowd after it has formed or try to guess the future based on simple counts of people, missing the complex web of connections that actually drives the growth.

A team of researchers has developed a new way to watch these events unfold, treating them not just as a list of attendees, but as a living, breathing network of relationships. They call their system auto-ibDLM, a tool designed to predict how a crowd will grow by studying the evolving patterns of how people talk to and influence one another online. Instead of trying to count every single person in real-time, which is often impossible, the system watches the structure of the conversation. It looks at how new people join the discussion, how existing participants link up, and how these connections change from hour to hour. By mapping these interactions, the researchers can forecast the next wave of arrivals with surprising precision, offering a window into the future of an event before it reaches a critical point.

The core idea behind this work is that the growth of a public event is fundamentally driven by the dynamics of its participants. When people share, comment on, or repost content about an event, they are not just passive observers; they are active nodes in a network that pulls others in. The researchers realized that if they could track how this network of interactions changes over time, they could predict how many new people would join the event in the next hour. To test this, they gathered data from thirteen real-world public events on a major Chinese social media platform, covering topics ranging from pandemic updates to local traffic jams and cultural festivals. They broke these events down into hourly snapshots, creating a timeline of how the network of participants grew and shifted.

To make sense of this complex data, the researchers built a framework that combines two different ways of looking at the world. First, they used established principles from network science to measure the shape of the crowd's interactions. They calculated things like how far apart people were in the network, how dense the connections were, and how groups of people were forming or breaking apart. These measurements provided a solid, understandable foundation. However, they knew that simply feeding these numbers into a computer would not be enough, because the relationship between these shapes and future growth is messy and changes from event to event. So, they added a second layer: an automatic learning system that could find the hidden patterns within those measurements. This system learned to compress the complex data into a simpler, more powerful form that highlighted the most important signals while ignoring the noise.

Once the system had this refined understanding of the network's current state, it used a specialized type of memory model to look at the sequence of changes. This allowed it to see not just what happened, but how the pattern of change was accelerating or slowing down. The result was a forecast of how many new participants would arrive in the next hour. When they tested this against other methods, including standard statistical tools and advanced artificial intelligence models, their system consistently outperformed them all. It was able to predict the size of the crowd with an accuracy that exceeded 97 percent in many cases, catching the sudden surges that other models missed or misjudged.

Perhaps the most valuable aspect of this discovery is that the system does not act as a black box. Because it starts with clear, human-understandable measurements of the network, the researchers can look inside and see which factors are driving the prediction. They found that certain types of connections, such as how tightly knit small groups were becoming or how the spread of information changed direction, were the strongest indicators of a coming surge. This transparency means that safety officials can not only get a warning but also understand why the warning was issued. For instance, the system might flag an event because the network structure shows that a few key individuals are beginning to pull in large clusters of new people, a sign that the event is about to outgrow its safe capacity.

The researchers also tested their method on other types of networks, such as online communities and digital currency exchanges, to see if the approach worked beyond public events. It did, proving that the strategy of combining structural measurements with automatic learning is a robust way to predict growth in any dynamic system. While the current model focuses on the structure of interactions rather than the specific words people use, the success of the approach suggests a clear path forward. By adding the meaning of the conversations to the structure of the connections, future versions could become even more accurate. For now, this work offers a powerful new lens for watching the pulse of public life, turning the chaotic evolution of crowds into a predictable pattern that can be managed before it becomes a crisis.

Technical Summary: A Network-driven Framework for Public Event Forecasting via Dynamic Interaction Network Evolution

1. Problem Statement

Public events, characterized by large numbers of participants in confined spaces and limited time windows, pose significant risks to public safety and urban infrastructure when they escalate rapidly (e.g., the Itaewon crowd surge). While existing research has made progress in post-event management, resource allocation, and situational awareness, there is a critical gap in anticipating the evolution of public events before risks materialize.

Current forecasting limitations stem from a lack of early warning capabilities. The paper posits that for many public events, evolution is fundamentally driven by participant dynamics. Adverse outcomes often occur when participation intensity exceeds safe capacity limits. Therefore, forecasting future participant growth serves as an effective surrogate for anticipating public event evolution.

In the digital era, participant interactions on online platforms (sharing, discussing, commenting) drive event dynamics. The authors propose modeling these events as dynamic interaction networks, where nodes represent participants and edges represent interactions. Consequently, the problem of public event forecasting is reformulated as forecasting the evolution of these dynamic interaction networks, specifically predicting future node increments (new participants).

2. Methodology: auto-ibDLM

The paper proposes auto-ibDLM (interaction behavior-driven deep learning model), a network-driven deep learning framework. The architecture consists of four stages:

2.1 Problem Modeling

An event is represented as a dynamic interaction network $G_e = (\Delta G_e^1, \Delta G_e^2, ..., \Delta G_e^t)$ , where each temporal slice contains new participants and new interactions. The forecasting task is defined as predicting the node increment $y_{t+1}$ (number of new participants at $t+1$ ) based on the most recent $k$ network snapshots. The event is identified as a potential public event if the predicted increment exceeds a predefined threshold $\delta$ .

2.2 Structural Representation (Hybrid Strategy)

Instead of using raw graph embeddings which may lack interpretability and incur high computational costs, the framework employs a hybrid representation learning strategy:

Network Science-Informed Structural Metrics: Each network snapshot is transformed into a structural feature vector comprising:
- Network-level metrics: Diameter, average shortest path length, and network density.
- Node-level metrics: 11 widely used metrics (e.g., Degree, Extended Degree, Clustering Coefficient, Coreness) computed for both incoming and outgoing connections. For each metric, the mean and standard deviation across all nodes are calculated to create a compact 44-dimensional vector.
- Community evolution behaviors: Six behaviors (Maintain, Dissolve, Merge, Split, Grow, Shrink) are tracked by matching communities across consecutive snapshots, yielding a 12-dimensional vector.
- The final structural feature vector for a transition is the concatenation of the temporal differences of network/node metrics and the community evolution vector.
Auto-Learning Layer: Recognizing that the relationship between handcrafted structural features and future growth is complex and non-stationary, an auto-learning layer (a fully connected linear transformation) projects the high-dimensional structural vectors into a compact and robust latent representation.
- To ensure the latent representation preserves predictive information, a consistency loss is introduced. This loss encourages the temporal variations of the latent representation to align with the temporal variations of the actual node increments, effectively filtering out non-informative features.

2.3 Temporal Forecasting

The learned latent representations are fed into a Gated Recurrent Unit (GRU) based temporal module. The GRU captures sequential dependencies in the latent space and, combined with historical node increments, predicts the future node increment via a fully connected layer.

3. Key Contributions

Novel Formulation: The paper explicitly formulates public event forecasting as an evolution forecasting problem on dynamic interaction networks, focusing on predicting node increments (participant growth) rather than raw event labels.
Hybrid Representation Learning: It introduces a strategy that combines network science-informed structural metrics with automatic representation learning. This approach reduces dependence on large-scale event-specific training data while providing representation-level interpretability by linking latent dimensions back to specific structural metrics.
Comprehensive Validation: The framework is validated on 13 real-world Sina Weibo public event datasets and two general dynamic network datasets (Superuser and Bitcoin), demonstrating superior performance over state-of-the-art methods.

4. Experimental Results

The framework was evaluated against sequence models (Transformer, LSTM, CNN-LSTM), dynamic graph representation learning models (EvolveGCN, WinGNN, SiGNN), and statistical models (VAR, ARIMA).

Forecasting Accuracy: On the Weibo benchmark, auto-ibDLM achieved high accuracy in public event identification across varying thresholds. Specifically, under the LPA community detection scheme, it attained 97.3% accuracy at a threshold of 200 participants, reaching 100% accuracy at thresholds of 300, 400, and 500 participants, and 97.4% at 600 participants. Under the LMC-APR scheme, it achieved 98.7% at 200 participants and maintained 100% accuracy across thresholds from 300 to 600. In terms of participant growth prediction (regression), the model consistently outperformed all baselines in Mean Squared Error (MSE), successfully capturing both overall evolution trends and sudden participant surges which other models often missed or overestimated.
Generalizability: The model maintained superior performance on the Superuser and Bitcoin datasets, which exhibit smoother and different evolution characteristics compared to the surge-driven Weibo data, indicating strong generalization capabilities.
Interpretability: Analysis of the auto-learning weight matrix revealed that specific structural metrics (e.g., standard deviation of in-COREDPA, out-EXTD) contributed most significantly to the latent representations. This aligns with domain knowledge while also capturing subtle patterns difficult to identify manually.
Ablation Studies: Removing the auto-learning layer or the consistency loss significantly degraded performance, confirming the necessity of the hybrid strategy. Furthermore, experiments showed that using only node-level structural metrics was sufficient to achieve excellent performance, suggesting a computationally efficient deployment option.

5. Significance and Claims

The paper claims that auto-ibDLM provides an effective and practical solution for intelligent public event forecasting. Its significance lies in:

Proactive Risk Management: By forecasting participant growth before capacity thresholds are exceeded, the system enables timely interventions such as access control or crowd diversion.
Bridging Knowledge and Data: The hybrid strategy successfully integrates network science knowledge (structural metrics) with data-driven deep learning, addressing the challenge of limited historical data for specific events while maintaining interpretability.
Robustness: The framework demonstrates robustness across diverse network types, suggesting its applicability beyond just public events to general dynamic network evolution forecasting.

The authors acknowledge limitations, such as the current focus on structural evolution without distinguishing interaction semantics (e.g., supportive vs. opposing opinions), and propose future work to integrate multi-modal information and advanced NLP techniques.