Network Adjusted Statistical Learning from Heterogeneous Crisis News Streams: An Acquisition Aware Framework for Crisis Informatics

Introduction:
Online crisis news streams help monitor disasters, infrastructure failures, transport disruptions, industrial incidents, and emergency events. However, raw counts can be distorted by acquisition volume, source coverage, temporal unevenness, and class imbalance. This study develops a network adjusted statistical learning framework for modelling heterogeneous crisis news as acquisition aware relational data rather than direct incident counts.

Methods:
The study analyses 71,930 crisis news records collected from 26 September 2023 to 2 June 2026, covering 20 event categories and 772 source domains. Records were cleaned, monthly indexed, and grouped into six monitoring periods, with P1, September 2023 to March 2024, used as the baseline period and P6, December 2025 to June 2026, used as the recent period. Event labels were derived from structured event codes and normalised into a curated 20 category crisis taxonomy. The framework integrates acquisition exposure normalisation, source event network modelling, event similarity estimation, imbalanced classification, feature ablation, and rare class transfer. Rare class transfer uses relational and semantic structure from better represented categories to improve low frequency classification. Classification results are reported from a stratified evaluation and interpreted cautiously, with source independent and temporal holdout validation identified as required robustness checks.

Results:
Raw crisis news volume declined between P1 and P6, but exposure adjustment showed that much of the apparent decline reflected reduced acquisition coverage rather than reduced reporting intensity. The taxonomy was highly imbalanced, with four dominant categories representing 67.11% of records. Classification performance was moderate but informative, with macro F1 = 0.564 and balanced accuracy = 0.559. Rare class transfer improved Significant Earthquake Event from 0.00 to 0.47 F1 and National Infrastructure Event from 0.00 to 0.19 F1.

Conclusions:
The framework supports acquisition corrected, network aware, and statistically cautious crisis informatics for event classification, source interpretation, open source crisis monitoring, and AI assisted situational awareness.