This paper presents an interesting, and socially valuable, real-world application report to the NeurIPS community. It provides a good balance of theoretical exploration and practical analysis, and the performance improvement on real-world data is compelling. In assessing "application reports" from all domains for suitability for NeurIPS, we often ask "what would the general ML practitioner learn from this application", and here I believe the learning about ML is relatively small, however I do believe that exposure to the problem has value to the community, so I lean on the side of acceptance.