Learning Large Neighborhood Search Policy for Integer Programming (Appendix) A.1 Architecture of bipartite GCN

In this paper, we propose to factorize the selection of a variable 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. MLP module that computes probabilities of selecting each variable in parallel.Figure A.1: Illustration of our LNS framework with the bipartite GCN based destroy operator. Our RL algorithm for training LNS policies is depicted by the pseudo code in Algorithm 1. The architecture of the neural network is displayed in the upper half of Figure A.2, which MLPs by a parameter-sharing MLP, as shown in the lower half of Figure A.2. S. / D. V ariable features ( V) Normalized reduced cost. 1 S. Normalized objective coefficient. 1 S. Normalized LP age. 1 S. Equality of solution value and lower bound, 0 or 1. 1 S. Equality of solution value and upper bound, 0 or 1 . 1 S. Fractionality of solution value. 1 S. One-hot encoding of simplex basis status (i.e., lower, basic, upper).