MixTTE: Multi-Level Mixture-of-Experts for Scalable and Adaptive Travel Time Estimation

TL;DR

MixTTE introduces a scalable, adaptive framework that plugs link-level spatio-temporal modeling into an industrial route-centric travel time estimation system. It combines a Spatio-Temporal External Attention (STEA) module for global road-network context with an Externally Stabilized Graph MoE (ESGMoE) to handle heterogeneous, long-tail traffic patterns, augmented by an Asynchronous Incremental Learning (ASIL) strategy for real-time adaptation. The approach yields consistent offline gains over multiple baselines and demonstrates robust online performance, including significant improvements in long-tail scenarios and successful production deployment at DiDi. This work demonstrates practical impact by delivering higher accuracy and stability for TTE in large-scale, dynamic urban networks, enabling more reliable ride-hailing services.

Abstract

Accurate Travel Time Estimation (TTE) is critical for ride-hailing platforms, where errors directly impact user experience and operational efficiency. While existing production systems excel at holistic route-level dependency modeling, they struggle to capture city-scale traffic dynamics and long-tail scenarios, leading to unreliable predictions in large urban networks. In this paper, we propose \model, a scalable and adaptive framework that synergistically integrates link-level modeling with industrial route-level TTE systems. Specifically, we propose a spatio-temporal external attention module to capture global traffic dynamic dependencies across million-scale road networks efficiently. Moreover, we construct a stabilized graph mixture-of-experts network to handle heterogeneous traffic patterns while maintaining inference efficiency. Furthermore, an asynchronous incremental learning strategy is tailored to enable real-time and stable adaptation to dynamic traffic distribution shifts. Experiments on real-world datasets validate MixTTE significantly reduces prediction errors compared to seven baselines. MixTTE has been deployed in DiDi, substantially improving the accuracy and stability of the TTE service.

Figures (7)

• Figure 1: The long-tail distributions of ride-hailing trip data w.r.t. three route-centric measurements. A state-of-the-art route-centric TTE method ieta2023kdd still performs poorly on a non-negligible fraction of tail routes.
• Figure 2: Overall framework of MixTTE. The left part illustrates the pipeline of multi-level TTE from the view of a specific road link. The right part depicts the asynchronous strategy at the start of each IL update step.
• Figure 3: Relative MAE gains of MixTTE over DiDi's current TTE model ieta2023kdd across head and tail traffic scenarios.
• Figure 4: Ablation study on Nanjing and Suzhou datasets.
• Figure 5: Interpretability analysis of ESGMoE layers in Beijing dataset on 2024-09-02. The color of each matrix element indicates the proportion of links assigned to a given expert that are experiencing congestion or non-recurring conditions at each time step. Here, the condition is defined as non-recurrent if the deviation between the current traffic condition and its historical average exceeds the 0.9 quantile of the overall historical distribution.

Theorems & Definitions (3)

• Definition 1: Traffic Network
• Definition 2: Traffic Slice
• Definition 3: Route