With the rise of online navigation and ride‑hailing services, optimal route planning that respects diverse user preferences has become a hot topic. Classic graph algorithms rely on heuristic cost functions to assign edge weights, offering no guarantee of route quality optimality. Existing data‑driven methods treat the ground‑truth optimal route as the observed user trajectory, but this is influenced by the navigation service and suffers from a feedback‑loop problem. To overcome these issues, we propose a deep architecture that jointly optimizes the cost function and the route‑ranking model to accommodate any route preference.
First, we run a multi‑objective Dijkstra algorithm offline to gather the set of Pareto‑optimal routes, treating it as the complete candidate set. Exploiting this property, we design a neural network that emulates shortest‑path search and route ranking in an end‑to‑end differentiable fashion.
Second, we formulate route preference as a constrained optimization task, enforcing constraints on route attributes while optimizing a single‑objective variable, and introduce a novel loss that reflects this formulation.
Extensive experiments on real‑world datasets show that our architecture markedly outperforms state‑of‑the‑art methods in both route quality and customizability.
Review