Self-Supervised Pre-trained Neural Network Models for Large-Scale Social Media Event Detection Tasks

Read full text Summarise with VIRAT
Authors: Ashu Nayak, Omprakash Dewangan
Cited by: 0
References: 40

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Social media platforms generate massive streams of multimodal data that can predict real-world events, such as political upheavals and natural disasters. Situational awareness, public safety, and decision-making require real-time detection. Traditional supervised event detection methods are limited by the brief, noisy, and multimodal nature of social media posts, which often have sparse labels and rapidly changing conditions. Large-scale manual annotation is impossible, and pre-trained techniques often overlook the social graph and temporal dynamics required for accurate detection. Self-supervised Pretraining for Event Detection (SPED), a multimodal approach that leverages unlabeled social media data to build robust representations, may address these issues. SPED integrates four complementary self-supervised goals: semantic grounding through masked language and image modeling, text-image coherence through cross-modal contrastive alignment, temporal progression prediction to capture event dynamics, and community propagation patterns through graph context modeling. The single backbone encodes text, image, temporal, and user-interaction signals, enabling adjustable fine-tuning for early detection, classification, and cross-platform transfer. SPED outperforms strong baselines in extensive datasets from Reddit and Twitter, reducing the average time-to-detection by 18% and increasing the F1-score by 12%. Ablation experiments confirm temporal and graph-aware targets’ early recall and robustness against noisy or incomplete data. In conclusion, SPED establishes a scalable, generalizable, and effective pretraining paradigm for social media event recognition, laying the groundwork for adaptive crisis monitoring systems.

Fields of Study & Subjects

Multimodal Machine Learning Applications Topic Modeling Public Relations and Crisis Communication

Publication details

Journal / source
Procedia Computer Science
Volume
275
Pages
242–249
Publication date
1 January 2026
Language
EN