Towards Explainable Neural Networks for Mathematical Programming