AI Applications in Mobility: Route Planning, ETA Prediction, Dynamic Event Mining, and Global Scheduling

Introduction: This article is the first in the #SpringRecruitmentSeries# organized by Amap (Gaode) Technology Tribune, summarizing the AI applications in the travel domain presented by Damon, head of the Amap Machine Learning R&D department.

AT Technology Tribune is a technical exchange program where experts discuss topics related to AI and travel services.

Pre‑travel – Route Planning: Route planning is treated as a recommendation problem similar to web or product search. The recommended route must be reachable, of high quality, and personalized ("千人千面"). Additional criteria include superiority and diversity compared with the primary route.

Characteristics of the planning algorithm: User queries contain implicit preferences such as personalization (familiar routes, preferences for time, distance, traffic lights, tolls) and scenario‑based needs. User ID stores familiar routes, while a DIN‑style network models preferences from historical navigation sequences.

The navigation engine performs recall, ranking, and filtering. Recall must be fast, ranking balances multiple objectives (e.g., travel time vs. user satisfaction), and filtering enforces constraints such as avoiding bad cases.

Multi‑objective optimization is modeled as a constrained optimization problem to improve specific metrics without degrading others.

Route Planning Recall Algorithms: Classical Dijkstra is too slow for long distances; A* improves speed but may lose optimality. Industry practice relies on cache‑based shortcuts. The Contraction Hierarchies (CH) algorithm preprocesses node importance and adds shortcuts for fast bidirectional queries.

Cache updates are costly for CH, leading to the next‑generation Cell‑Based Routing (CBR). CBR partitions the road network into hierarchical sub‑graphs, precomputes shortcuts between boundary nodes, and stores them in CPU L1 cache, achieving update times of ~15 seconds for the national network.

Comparison of CH and CBR: Query latency: CH ~0.1 ms, CBR 1–2 ms. Shortcut update: CH ~10 min, CBR ~15 seconds. CH shortcuts are irregular, causing high memory usage for different routing strategies; CBR shortcuts are more reusable.

Sorting Network: A bias network mitigates the strong preference for the top‑ranked route (over 90% selection). The network incorporates user bias, historical statistics, and navigation sequence features to improve actual travel coverage.

Pre‑travel – ETA Prediction: Predicting future traffic conditions relies on spatio‑temporal modeling. Temporal patterns are captured with RNN/LSTM/Seq2Seq, while spatial relationships use Graph Neural Networks. Historical data alone struggles with sudden events (e.g., large gatherings).

To handle unexpected congestion, the volume of planned trips is fed into the traffic‑prediction model, improving accuracy during spikes. This approach was accepted at KDD 2020.

In‑travel – Dynamic Traffic Event Mining: Unstructured text from media reports is processed to extract incident elements. BERT models are used for feature extraction, but their latency is high for online use.

Knowledge Distillation: A lightweight student network is trained via contrastive learning‑based knowledge distillation, using cosine distance instead of Euclidean distance. This method was published at AAAI 2021.

Post‑travel – POI Data Freshness: Enhancing the timeliness of Point‑of‑Interest data involves quickly detecting closed or moved venues and marking them offline. Reference: "Gaode POI Freshness Enhancement".

Future Outlook – Global Scheduling: Unlike search results, road resources are limited. A global scheduling system integrates with traffic signals and employs multi‑agent reinforcement learning to coordinate vehicles and traffic lights, aiming to improve road utilization and reduce congestion.

References: KDD 2020 paper on hybrid spatio‑temporal graph convolutional networks, AAAI 2021 paper on LRC‑BERT, and Gaode’s 2020 technology collection ebook.