Breaking the Black Box: A New View of Fake News Detection Using Multi-Level TrustScore and Micro-Trees

Abstract

Rumor detection on social media remains a critical challenge due to the unstructured and hierarchical nature of conversational threads. While deep learning models often act as black boxes and require massive computational resources, traditional Machine Learning (ML) offers native interpretability but suffers from flat feature representations. In this paper, we introduce a novel feature engineering framework that transforms each individual tweet into a fine-grained Micro-Tree topology (an intra-tweet content dependency graph) and extracts a deterministic multi-level numerical vector, the TrustScore, via a breadth-first traversal across propagation levels. A forward-fill imputation policy allows reduction of the full 47-level feature space to 4 statistically validated variables (Mann-Whitney U test, p < 10^−18): User Health, Source Tweet Objectivity, Effective Depth, and the inherited terminal score Score L46. After targeted data cleaning (two pathological events excluded), we evaluate five classifier families on 7 PHEME events (N = 6,178 threads) under a strict Leave-One-Event-Out (LOEO) protocol, which eliminates cross-event data leakage inherent in standard k-fold evaluation. Random Forest achieves the best Macro-F1 (0.5722) while XGBoost reaches the highest global accuracy (0.6141). Decision threshold tuning (θ = 0.10) achieves rumor recall exceeding 99.8%, enabling early-warning deployment. The framework is CPU-efficient (avg. 1.27 s/fold), fully interpretable by design, and provides concrete step-by-step explainability without reliance on post-hoc tools such as SHAP or LIME.