QMF-TEA: a random-forest-based data-quality-aware multi-feature fusion method for trajectory entity alignment in road freight transportation

Multi-source trajectory records in road freight transportation often differ in schema, representation, sampling interval, field availability, and data quality, making records of the same trip difficult to link reliably. This study proposes QMF-TEA, a random-forest-based data-quality-aware multi-feature fusion method for trajectory entity alignment under weak-identifier conditions. QMF-TEA constructs cross-view candidate pairs and integrates spatial, road-network, trajectory-morphology, sampling-interval, source-view, map-matching-error, and quality features, while excluding strong identifiers and other leakage-prone fields. On 32,436 candidate pairs, the primary held-out test achieved a precision of 0.8842, recall of 0.9646, F1-score of 0.9226, and AUC of 0.9772. Complementary group-disjoint evaluations yielded balanced F1-scores of 0.9268 ± 0.0065 for unseen entities and 0.9199 ± 0.0090 for unseen vehicles, supporting generalization beyond pair-level random splitting. Repeated paired ablation showed that quality features improved F1-score by 0.0067 and AUC by 0.0077, with both gains significant under exact two-sided sign-flip tests ( p = 0.0020). Spatial and road-network features remained the dominant evidence, whereas source and sampling features provided auxiliary context. A separately validated clustering threshold of 0.85 and conflict-aware merging were used to control non-transitive pairwise links. Finally, reference-group analysis quantified the upper-bound potential for duplicate discovery, field completion, quality screening, and conflict discovery; these values are not presented as end-to-end governance outcomes from predicted clusters. The results demonstrate an interpretable and leakage-aware approach to multi-view trajectory alignment and provide a defensible basis for subsequent road-freight data governance.