Freeway Traffic Incident Duration Prediction via Multi-Source Semantic Fusion and a Mixture-of-Experts

Abstract

Most incident duration models emit only a point estimate and are evaluated on archives whose recorded durations partly reflect logging conventions. We propose a mixture-of-experts (MoE) model that fuses structured attributes, incident-text semantics, and environmental context, restricted to inputs available at the first report, and that delivers point predictions, duration classes, and calibrated quantile intervals from a single inspectable model, with monotone quantiles by construction. A label-quality and construct-validity audit is equally central: about 15% of the recorded durations in the public US Accidents archive (400,001 records, 2016–2023) carry source-specific default or quantized signatures, and re-training on default-free labels shows that apparent point-accuracy advantages partly reflect that label structure. On the 2022–2023 test window, the five-seed MoE ensemble reaches 58.54 min mean absolute error versus 61.52 min for an identically configured gradient-boosted baseline (paired difference 2.98 min, 95% CI [2.66, 3.29]), and is statistically equivalent within a pre-specified ±1.0 min margin on default-free labels (61.26 versus 61.55 min). Its 80% intervals cover 0.793 with zero crossings, whereas three separate tree models under-cover by ten points, and text lowers the error by 5.21 min (3.50 min when timing-unverifiable phrases are masked). A bootstrap shows negligible test-reuse bias, and two temporal re-selections that never consult the test window show that the text gain and the margin over the trees hold on records without congestion-state phrasing (54–62% of each window) and fail on records with it, whose relation to duration changed in mid-2021.