ARCHIVES

The accelerating pace at which topics rise, peak, and fade across digital platforms has made early trend detection a critical capability for journalism, public health surveillance, marketing, and policy planning. Conventional trend-detection systems rely on single-platform signals and opaque deep learning models, which limits both their coverage and their trustworthiness in operational decision-making contexts. This paper proposes a multi-source social media analytics framework that fuses textual, temporal, and network signals from Twitter/X, Reddit, YouTube comments, and news RSS feeds to detect emerging trends substantially earlier than single-source baselines. The framework combines a burst-aware temporal feature extractor, a graph-based diffusion model capturing cross-platform propagation, and a gradient-boosted forecasting ensemble, with SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) layers providing per-prediction interpretability. We evaluate the framework on a curated 14-month, multi-platform dataset spanning approximately 9.2 million posts across 1,800 manually and weakly labeled emerging topics. The proposed system achieves a mean detection lead time of 36.4 hours ahead of trend-platform native trending lists, an F1-score of 0.91 for trend/non-trend classification, and a mean absolute percentage error (MAPE) of 11.3% for 48-hour volume forecasting, outperforming LSTM, Prophet, and single-platform XGBoost baselines. Explainability analysis shows that burst acceleration, cross-platform co-occurrence, and influencer-adjusted reach are the most consistently influential features across topic categories. The results demonstrate that combining multi-source fusion with explainable machine learning yields both earlier and more trustworthy trend forecasts, with direct implications for misinformation early-warning systems, public health surveillance, and real-time marketing intelligence.