UNITE: A Unified Framework for Accurate and Efficient Origin-Destination and Route Travel Time Estimation

Travel Time Estimation (TTE) on road networks is crucial for modern location-based services. There are two main variants: Origin-Destination TTE (ODTTE), which estimates travel time between an origin and destination given a departure time, and Route TTE (RTTE), which estimates travel time along a specified route. Existing approaches typically address ODTTE and RTTE separately. We propose UNITE, a unified framework that efficiently addresses both ODTTE and RTTE with a single training process. UNITE operates in a progressive, segment-wise manner. We first design a Progress State Encoder (PSE) that produces a shared state embedding capturing the origin, destination, departure-time context, sequential dependencies, and dynamic traffic conditions along the partially generated route. Specifically, PSE constructs segment-level travel-time histograms from historical trajectories and applies attention-based sequence modeling. Next, we develop a Segment Travel Time Estimator (mSTE) that leverages the shared state embedding, builds a local spatio-temporal travel graph to capture traffic dynamics, and employs a Mixture-of-Experts architecture to model heterogeneous traffic patterns for segment-level travel time estimation. We further introduce a Next Segment Prediction (NSP) module that predicts the next segment from the outgoing neighbors of the current one based on the shared state embedding, incorporating multi-hop look-ahead neighborhoods. Extensive experiments on large-scale real-world datasets demonstrate that UNITE consistently outperforms state-of-the-art methods in both accuracy and efficiency.