This paper provides contributions regarding generalization in the predict-then-optimize framework, that suggests to value the quality of a ML system not only wrt to some prediction loss but to a decision loss. The contribution is of quality, addressing the little studied SPO framework in ML community. In addition to the theoretical results, it would be nice if the authors could squeeze in more examples (in the line of the shortest path problem), this would contribute to an effort aiming at providing some perspective on the paper.