GenMRP: A Generative Multi-Route Planning Framework for Efficient and Personalized Real-Time Industrial Navigation

Existing industrial-scale navigation applications contend with massive road networks, typically employing two main categories of approaches for route planning. The first relies on precomputed road costs for optimal routing and heuristic algorithms for generating alternatives, while the second, generative methods, has recently gained significant attention. However, the former struggles with personalization and route diversity, while the latter fails to meet the efficiency requirements of large-scale real-time scenarios. To address these limitations, we propose GenMRP, a generative framework for multi-route planning. To ensure generation efficiency, GenMRP first introduces a skeleton-to-capillary approach that dynamically constructs a relevant sub-network significantly smaller than the full road network. Within this sub-network, routes are generated iteratively. The first iteration identifies the optimal route, while the subsequent ones generate alternatives that balance quality and diversity using the newly proposed correctional boosting approach. Each iteration incorporates road features, user historical sequences, and previously generated routes into a Link Cost Model to update road costs, followed by route generation using the Dijkstra algorithm. Extensive experiments show that GenMRP achieves state-of-the-art performance with high efficiency in both offline and online environments. To facilitate further research, we have publicly released the training and evaluation dataset. GenMRP has been fully deployed in a real-world navigation app, demonstrating its effectiveness and benefits.

💡 Research Summary

The paper introduces GenMRP, a generative multi‑route planning framework designed for large‑scale industrial navigation where real‑time, personalized, and diverse route recommendations are required. Existing solutions fall into two categories: (1) pre‑computed link cost methods that excel at finding the optimal path but struggle with personalization and diversity, and (2) recent generative approaches that can model user preferences but are too computationally heavy for industrial‑scale, low‑latency deployment. GenMRP bridges this gap with four key innovations.

First, the “Skeleton‑to‑Capillary” (STC) technique constructs a request‑specific sub‑network that is orders of magnitude smaller than the full road graph. By extracting a skeleton of critical links (e.g., those near the origin, destination, and major intersections) and then expanding locally with capillary links that preserve dynamic traffic, turn‑restriction, and contextual information, the system reduces memory and CPU usage while retaining the essential structure needed for accurate routing.

Second, a Link Cost Model (LCM) predicts per‑link costs using a combination of static attributes (length, road type, tolls), dynamic attributes (real‑time speed, congestion, weather), user historical sequences, and previously generated routes. The LCM leverages Deep Interest Network (DIN) to capture user‑link interactions, Gated Attention Transformers (GAT) for neighbor‑link relationships, and a Multi‑Scenario Network to adapt to different contexts (e.g., commuting vs. long‑distance trips). The model outputs asymmetric, context‑aware costs that are fed into the routing engine each iteration.

Third, route generation is performed iteratively with a bidirectional Dijkstra (BDR) module. In the first iteration the algorithm returns the route with the highest coverage of the user’s actual trajectory (the metric used throughout the paper). Subsequent iterations employ a novel “correctional boosting” objective: the system maximizes the increase in coverage (ΔCov) relative to the set of routes already produced, while still encouraging high absolute coverage. This formulation naturally balances diversity (by penalizing overlap with earlier routes) and quality (by rewarding routes that still align well with the user’s preferences).

Because BDR is non‑differentiable, direct gradient‑based training of the LCM is impossible. The authors adopt a sampling‑based learning scheme: multiple candidate routes are generated using multi‑dimensional Pareto link costs; the candidate with the highest trajectory coverage becomes a positive sample, and the rest become negatives. The LCM is then trained with a binary cross‑entropy loss to increase the probability of generating high‑coverage routes. This indirect supervision aligns the model’s objective with the ultimate evaluation metric.

A major contribution is the release of the PersonaRoadNet (PRN) dataset, collected from a real‑world navigation app. PRN contains about 590 k unique user requests, each with an average of 1 300 links, 200+ dynamic and static features per link, user familiarity frequencies, heat‑map statistics, and full historical sequences. This request‑level dataset is far richer than publicly available OSM or Natural Earth data and enables reproducible research on personalized, diverse, and real‑time routing.

Extensive experiments validate GenMRP. Offline benchmarks against k‑shortest path, penalty‑based alternatives, and recent generative baselines show improvements of over 10 % in coverage and 15 % in diversity metrics (e.g., Jaccard similarity, NDCG). The average inference latency, including sub‑network extraction, LCM update, and Dijkstra search, stays below 30 ms, meeting real‑time constraints. Online A/B testing on a production navigation service (over 1 M users for two months) reports a 12 % lift in the proportion of recommended routes that users actually follow, and a 38 % reduction in CPU load, confirming both user‑experience and system‑efficiency gains.

In summary, GenMRP contributes (1) an efficient sub‑graph construction method (STC), (2) a context‑aware, user‑personalized link cost model, (3) a correctional boosting mechanism that jointly optimizes quality and diversity across multiple iterations, (4) a practical training pipeline for non‑differentiable routing, and (5) a comprehensive, publicly released dataset. These advances collectively set a new standard for real‑time, large‑scale industrial navigation systems.