Source-linked AI summary
A Network-driven Framework for Public Event Forecasting via Dynamic Interaction Network Evolution
Jie Wei, Yue Liu, Xiaochuan Tang, Biao Cai, Xiangtao Li, Yanmei Hu
TL;DR
Forecasting public event evolution is difficult because participant interactions dynamically drive participation and can create safety risks. auto-ibDLM models events as evolving interaction networks, learns structural representations, and forecasts participant growth, outperforming existing methods across public-event datasets and generalizing across dynamic networks.
Problem
Public event forecasting needs to anticipate participant growth driven by dynamic interactions so authorities can mitigate safety risks and allocate resources proactively.
Method
auto-ibDLM represents events as dynamic interaction networks, transforms structural metrics into latent representations, and uses a GRU module to forecast participant growth.
Results
auto-ibDLM consistently outperformed existing methods on 13 public-event datasets and generalized across dynamic networks with different evolution characteristics.
Takeaways & Limitations
Dynamic interaction network evolution provides an effective basis for public event forecasting, with node-level structural metrics offering an efficient deployment configuration.
Takeaways & Limitations
The framework models interaction structure without distinguishing semantic content, and further work is needed to exploit complementary information among structural metric categories.
Abstract
from arXiv · showhide
Effective public event forecasting is essential for intelligent service systems, enabling proactive risk management, adaptive resource allocation, and timely decision-making. In many real-world scenarios, the evolution of public events is driven by dynamic interactions among participants. Motivated by this observation, this paper proposes auto-ibDLM, a network-driven deep learning framework that represents events as dynamic interaction networks and predicts public event evolution through participant growth forecasting. The proposed framework adopts a hybrid representation learning strategy that first represents network evolution using network science-informed structural metrics and subsequently transforms the resulting structural feature vectors into compact and robust latent representations through an auto-learning layer. A GRU-based temporal forecasting module is then employed to capture temporal dependencies and predict future participant growth. Extensive experiments on 13 real-world public event datasets and two publicly available dynamic network datasets demonstrate that auto-ibDLM consistently outperforms representative state-of-the-art methods in both forecasting accuracy and generalization capability, achieving over 97% accuracy in public event forecasting. Comprehensive experimental analyses further validate the effectiveness of the proposed hybrid representation learning strategy and demonstrate its representation-level interpretability. These results indicate that auto-ibDLM provides an effective and practical solution for intelligent public event forecasting.
1 Introduction
Public event forecasting remains limited because existing work largely manages ongoing events rather than anticipating their evolution. The paper addresses this gap by modeling events as dynamic participant-interaction networks and forecasting future participant growth with auto-ibDLM.
- Motivation: Rapid escalation of large public events can pressure urban infrastructure and public services and threaten public safety.The introduction cites the Itaewon Halloween crowd surge, which caused 159 fatalities and 196 injuries.
- Motivation: Existing public-event research supports resource allocation, disaster assessment, situational awareness, and opinion management but largely responds to ongoing events.Consequently, the ability to forecast public event evolution before risks materialize remains limited.
- Motivation: Participant dynamics drive event evolution, and risks can arise when participation intensity exceeds safe capacity within a short period.Forecasting participant growth beyond a venue’s safe capacity can help organizers and emergency managers identify high-risk events.
- Network-centric formulation: Online sharing, discussion, and commenting both reflect public attention and influence participation, making interaction-network evolution a basis for forecasting event evolution.The paper formulates public event forecasting as forecasting the evolution of dynamic interaction networks.
- Proposed framework: auto-ibDLM represents events as dynamic interaction networks and forecasts future participant growth as node increments in the evolving network.The framework is presented as a network-driven deep learning model for public event forecasting.
2 Related Work
Related work spans public event forecasting and dynamic network evolution forecasting. Existing studies emphasize post-event analysis and embedding-based network prediction, while participant interactions and interpretable network-scale growth forecasting remain comparatively underexplored.
- Public event forecasting: Public event forecasting research has mainly analyzed social-media information and public responses after events, including disaster-relief requests, flood severity, and flood economics.These studies supported relief-supply distribution and post-event assessment but did not primarily forecast event development.
- Public event forecasting: Real-time social-media data can reveal emerging public events, with prior work detecting floods and analyzing public opinion and investor discussions.This line of work addresses early detection more directly than post-event response, but the passage does not establish participant-interaction modeling.
- Public event forecasting: Existing studies largely overlook participant interactions, although individual behaviors and interactions continuously shape participant-driven event dynamics and development.The paper therefore represents such events as dynamic interaction networks built from observable participant behaviors.
- Dynamic network evolution forecasting: Dynamic network evolution is studied at micro, meso, and macro levels, covering individual changes, substructure dynamics, and global node or edge growth.This taxonomy distinguishes interaction formation, community evolution, and network-scale growth over time.
- Dynamic network evolution forecasting: State-of-the-art methods predominantly use GNN-based network embeddings for dynamic prediction, but network-scale node-growth forecasting remains underexplored and often costly and difficult to interpret.The cited limitations concern substantial computational and memory overhead on large dynamic networks and limited representation-level interpretability.
3 Methodology
auto-ibDLM models public events as evolving dynamic interaction networks and forecasts future participant growth. It combines network structural and community-evolution features, auto-learned latent representations, and GRU-based temporal forecasting.
- Overall architecture: The four-stage framework models an event as a dynamic interaction network, extracts structural representations, learns latent features, and forecasts participant growth temporally.The stages are problem modeling, structural representation, auto-learning representation, and GRU-based temporal forecasting.
- Problem formulation: Participant growth is forecast as node increments, with predicted increment ŷt+1 determining public-event evolution when it exceeds threshold δ.The node increment is yt+1 = |∆Vt+1|, representing new participants joining at time step t + 1.
- Structural representation: Each snapshot is represented using network-level metrics, a 44-dimensional node-level vector, and a 12-dimensional community-evolution vector.Network-level metrics include diameter, average shortest path length, and network density; community behaviors include Maintain, Dissolve, Merge, Split, Grow, and Shrink.
- Auto-learning representation: An auto-learning layer automatically transforms structural feature vectors into compact and robust latent representations instead of relying on manually selected features.The layer addresses correlations between structural features and participant growth and variation in feature relevance across public events.
- Temporal forecasting: Stacked GRU layers encode the input sequence, concatenate final-layer hidden states from k − 1 units, and map them through a fully connected layer to predict ŷt+1.The GRU module captures temporal dependencies in recent network snapshots for participant-growth forecasting.
4 Experiments
Section 4 evaluates auto-ibDLM through public event forecasting, dynamic-network generalization, and ablation analyses. The experiments include a real-world Weibo benchmark and two publicly available dynamic network datasets.
- Experimental design: The experiments assess auto-ibDLM from three perspectives: public event forecasting, dynamic-network generalizability, and ablation analysis.The ablation study examines the hybrid representation learning strategy and different categories of structural metrics.
- Public event forecasting: Public event forecasting is evaluated using a real-world Weibo benchmark dataset.The supplied table summarizes the public events included in the Weibo benchmark dataset.
- Generalizability: Generalizability is evaluated by predicting node increments on two publicly available dynamic network datasets.
4.1 Experimental settings
The evaluation uses a unified benchmark of 13 Sina Weibo public events, with hourly dynamic-network sequences split chronologically for forecasting. Generalizability is additionally assessed on Superuser and Bitcoin, against sequence, dynamic-graph, and statistical time-series baselines under fixed training settings.
- Datasets: 13 Sina Weibo hot topics form the public-event benchmark, covering public concerns, social incidents, and tourism congestion events through participant interaction networks.Interactions include posting, commenting, liking, and reposting.
- Data construction: Hourly snapshots represent each event as structural-feature and node-increment sequences, using nine consecutive pairs with eight inputs and one prediction target.All topics were combined into a unified benchmark because each event provided limited samples.
- Data splitting: Chronological splitting assigns the first 90% of each topic to training and the remaining 10% to testing before merging topics into global sets.This procedure prevents temporal information leakage.
- Datasets: Superuser and Bitcoin provide complementary dynamic-network benchmarks for evaluating generalizability beyond public-event forecasting.Superuser contains 194,085 nodes and 1,443,339 edges with daily snapshots, while Bitcoin contains 5,881 nodes and 35,592 edges with weekly snapshots.
- Baselines: Comparisons cover sequence models, dynamic graph representation learning models, and statistical time-series models, including Transformer, EvolveGCN, VAR, and ARIMA.Sequence models use structural-feature sequences; dynamic graph models use historical edge lists; VAR uses multivariate features, whereas ARIMA uses node increments.
- Implementation: All models were trained for 3000 epochs with Adam at learning rate 10−3, while selected graph models used early stopping with patience 50.For auto-ibDLM, the auto-learning dimension was d′ = 10 and the prediction module used eight hidden layers with seven GRU cells and ten channels.
4.2 Public Event Forecasting
On the Weibo benchmark, auto-ibDLM most accurately forecasts participant growth, capturing both overall evolution and sudden surges. Thresholding predicted growth enables reliable public-event identification, while structural-metric contributions provide representation-level interpretability.
- Participant Growth Forecasting: auto-ibDLM produces the predictions closest to ground truth across both LPA and LMC–APR community detection schemes on Weibo.The test samples use sliding windows of nine consecutive network snapshots, and the target is participant growth, or node increments.
- Participant Growth Forecasting: auto-ibDLM most accurately forecasts participant growth, simultaneously capturing the overall evolution trend and sudden participant surges.Sequence, statistical, and dynamic graph representation learning baselines frequently miss surges, overestimate them, or underestimate growth.
- Representation Interpretability: The auto-learning weight matrix quantifies structural-metric contributions through mean absolute row weights, with higher scores indicating greater contribution to learned latent representations.The top three metrics are the standard deviations of in-COREDPA, out-EXTD, and out-COREDP.
- Representation Interpretability: Eleven of twelve highly ranked structural metrics coincide with manually identified informative metrics, indicating that the latent representations largely preserve domain-consistent structural characteristics.The standard deviation of in-COREDPA is the only exception and receives the highest contribution score.
- Representation Interpretability: Mean squared-weight analysis yields a ranking highly consistent with mean absolute weights, demonstrating the stability of learned structural contributions.Contribution scores reflect average contributions to latent representations, while redundancy and metric interactions can compensate for removed structural information.
- Public Event Identification: auto-ibDLM identifies potential public events by comparing predicted participant growth with thresholds of 200, 300, 400, 500, and 600 participants.It achieves nearly perfect recall across almost all thresholds, except 0 at 600 participants under LPA, where only two positive test samples exist.
4.3 Node increment prediction on general dynamic networks
auto-ibDLM consistently delivers the lowest node-increment prediction errors across Weibo, Superuser, and Bitcoin under both community detection schemes. Its advantage extends across sequence and dynamic graph representation learning baselines and is supported by temporal prediction curves and the hybrid representation strategy.
- Overall performance: auto-ibDLM achieves the lowest MSEs across Weibo, Superuser, and Bitcoin under both LPA and LMC-APR.On Superuser, it obtains MSEs of 0.000094 (LPA) and 0.000085 (LMC-APR).
- Sequence-model comparison: 0.000114 (0.000111) and 0.000154 (0.000100) are the Superuser MSEs achieved by LSTM and Transformer under LPA (LMC-APR), respectively.Their errors increase considerably on Weibo and Bitcoin, where node increments exhibit stronger fluctuations.
- Dynamic graph comparison: WinGNN consistently outperforms EvolveGCN-O and SiGNN, while EvolveGCN-H surpasses EvolveGCN-O on Weibo and Bitcoin and performs comparably on Superuser.None of these dynamic graph representation learning methods matches auto-ibDLM’s prediction accuracy.
- Temporal tracking: auto-ibDLM more accurately tracks the temporal evolution of node increments than competing methods across both general dynamic network datasets.The supplementary prediction curves support the quantitative MSE comparisons.
- Representation strategy: The hybrid representation learning strategy combines network science-informed structural metrics with automatic representation learning to produce robust latent representations across event-driven and general dynamic networks.This strategy underpins auto-ibDLM’s effectiveness and generalization across networks with substantially different evolution characteristics.
4.4 Ablation study
The ablation study shows that auto-ibDLM’s hybrid representation learning strategy is essential, with both the auto-learning layer and consistency loss contributing to superior forecasting. Fine-grained community metrics further help under LMC-APR, while node-level metrics alone remain highly effective.
- Ablation design: Six variants were constructed to assess the auto-learning layer, consistency loss, manually selected metrics, and structural metric configurations.The variants include ibDLM, auto-ibDLM−lossf, and manual-ibDLM.
- Representation learning: Across all datasets, auto-ibDLM consistently outperformed ibDLM, auto-ibDLM−lossf, and manual-ibDLM.This supports automatically transforming handcrafted structural feature vectors into compact latent representations.
- Representation learning: Both the auto-learning layer and the proposed consistency loss are indispensable components of auto-ibDLM.Neither ibDLM nor auto-ibDLM−lossf consistently matched auto-ibDLM across datasets.
- Structural metric configurations: auto-ibDLMc+l consistently outperformed auto-ibDLMl and auto-ibDLMm+l under LMC-APR, while the three variants achieved comparable performance under LPA.Fine-grained community evolution behaviors help when high-quality community detection results are available.
- Structural metric configurations: auto-ibDLMl achieved lower prediction errors than all competing methods in Table 6, showing that node-level structural metrics alone can yield excellent forecasting performance.This result demonstrates that network-level structural metrics are not required for strong forecasting performance.
5 Conclusion
auto-ibDLM frames public event forecasting as dynamic interaction network evolution forecasting through hybrid structural and automatic representation learning. The conclusion reports strong forecasting effectiveness, generalization, interpretability, and practical value, while identifying interaction semantics and adaptive representation learning as future directions.
- Contributions: auto-ibDLM represents public events as dynamic interaction networks and forecasts their evolution through future participant growth.Its hybrid representation learning combines network science-informed structural metrics with automatic representation learning.
- Findings: Modeling public event forecasting as dynamic interaction network evolution provides an effective basis for forecasting public event evolution.The framework consistently outperformed existing methods, while learned latent representations were primarily driven by node-level structural metrics and captured informative structural patterns.
- Findings: Evaluations on general dynamic networks show that auto-ibDLM generalizes across networks with substantially different evolution characteristics.Ablation results found that multi-level structural metrics performed best overall, whereas node-level metrics alone still achieved excellent forecasting performance.
- Limitations: The current framework models participant interactions through structural evolution without distinguishing semantic differences such as supportive versus opposing opinions.Incorporating semantic information may improve estimation of actual participant behavior and event forecasting.
- Future Work: Future work will integrate multi-modal information and advanced natural language processing to jointly model interaction structure and semantics.The authors also plan more adaptive representation learning to identify informative structural metrics and suppress redundant information.