Despite rapid progress in neural combinatorial optimization (NCO) for solving CO problems (COPs), as the problem scale grows, several bottlenecks persist: 1) solvers in the Global Prediction (GP) paradigm struggle in long-range decisions where the overly smooth intermediate heatmaps impede effective decoding, and 2) solvers in the Local Construction (LC) paradigm are time-consuming and incapable of tackling large instances due to the onerous auto-regressive process. Observing these challenges, we propose a new paradigm named Adaptive Expansion AE with its instantiation COExpander, positioned to leverage both advantages of GP and LC. COExpander utilizes informative heatmaps generated by a global predictor, which is learned under the guidance of locally determined partial solutions, to in turn direct the expansion of determined decision variables with adaptive step-sizes. To ensure transparent evaluation, we further take the lead to canonicalize 29 benchmarks spanning 6 popular COPs (MIS, MCl, MVC, MCut, TSP, ATSP) and various scales (50-10K nodes), upon which experiments demonstrate concrete SOTA performance of COExpander over these tasks. Source code and our standardized datasets will be made public.

We define the process of using machine learning to solve combinatorial optimization problems (COPs) as progressively determining decision variables. Previous approaches tend to be extreme: they either attempt to predict all decision variables in one shot (global prediction, GP) or determine them one by one in a sequential manner (local construction, LC). Through experiments, we find that GP solvers often lead to conflicts between predicted variables, while LC solvers is too inefficient. Therefore, we propose a new paradigm, namely adaptive expansion (AE), along with a supervised learning-based instance called COExpander. Compared to the LC solvers, COExpander requires fewer iterations. In each iteration, it adaptively selects a subset of variables using a global predictor, while the remaining variables are deferred for future inference. We evaluate COExpander on six classic COPs across 29 benchmarks and achieve SOTA results.