Beyond Engagement: Classifying News Impact on X Using Impressions Data and Hybrid Feature Architecture

Abstract

Predicting news reach on X (formerly Twitter) is a critical challenge for digital journalism. However, existing literature often relies on biased ``proxy metrics'' (likes, retweets) that miss the ``silent majority'' of passive viewers. This study presents an empirically validated multimodal framework for objective news impact classification using Impressions (View Count) as ground truth. We construct a time-stratified dataset from @detikcom on X (2022--2025), labeled into ”Regular,''  “Hot,'' and ”Viral'' classes via a data-driven log-base-10 thresholding strategy. A six-scenario ablation study evaluates our proposed Adaptive Gated Cross-Modal Fusion architecture against static baselines, contextual extractors, and standard MLP fusion variants. This architecture learns to dynamically balance textual and engagement signals via a trainable gating mechanism. The proposed model (S6) achieves a Macro F1-Score of 0.7383 (peak Viral-class F1 of 0.6277), representing an 85.8% improvement over the unimodal text-only baseline. A dual-track IndoBERT vs. mBERT comparison confirms the necessity of monolingual pre-training. Furthermore, forensic error analysis demonstrates how “semantic awareness'' resolves Cold Start underestimation and False Viral overestimation that consistently cause tree-based classifiers to fail.