LearningLargeNeighborhoodSearchPolicyfor IntegerProgramming(Appendix) A.1 ArchitectureofbipartiteGCN

In this paper,we propose to factorize the selection of avariable subset into decisions on selection of each variable, under our LNS framework. To represent such action factorization, we employ the bipartite GCN as the destroy operator, as shown in Figure A.1. Specifically,givenacollection of features for each variable, we first process them by a MLP to obtain variable embeddings. In this paper, we represent the state by a bipartite graphG =(V,C,A)attached by the features of variables, constraints andedges(i.e.V,CandA),which arelistedinTableA.1. Forthestatic features, weconsider theones used in[20], which learns variable selection policies in B&B algorithm.