Machine Learning for Two-Stage Graph Sparsification for the Travelling Salesman Problem

TL;DR

This paper introduces a two-stage machine learning-based graph sparsification method for TSP solvers, improving efficiency and generalization across various instance types and sizes.

Contribution

It proposes a novel two-stage approach combining heuristic union and learned reduction, outperforming existing neural methods especially at larger scales.

Findings

Significantly reduces candidate graph density while maintaining high coverage.

Outperforms recent neural sparsification methods across multiple distance types.

Generalizes well across different distributions and problem sizes.

Abstract

High-performance TSP solvers like LKH search within a sparsified candidate graph rather than over all possible edges. Graph sparsification is non-trivial: keep too many edges and the solver wastes time; cut too many and it loses edges that belong to the optimal tour. The two leading heuristic methods, $\alpha$-Nearest and POPMUSIC, produce high-quality candidate graphs, but no single heuristic is both sparse and reliable across all instance sizes and distributions. Machine learning methods can potentially learn better sparsification models. However, existing approaches operate on the complete graph, which is expensive and mostly restricted to Euclidean distances. To address this issue, we propose a two-stage graph sparsification approach: Stage~1 takes the union of $\alpha$-Nearest and POPMUSIC to maximise recall; Stage~2 trains a single model to reduce density. We conducted experiments across four TSPLIB distance types, five spatial distributions, and problem sizes from 50 to 500. The two-stage approach substantially reduces candidate-graph density while retaining high coverage, generalises across distance types and distributions, outperforms recent neural sparsification methods that are restricted to Euclidean distances, and becomes increasingly valuable at larger scales where single-stage heuristics degrade.