> Therefore, one often follows an alternative approach in practice, namely, to simply join the data of different tables, materialize the output of the multi-way

> Therefore, one often follows an alternative approach in practice, namely, to simply join the data of different tables, materialize the output of the multi-way join in one large table, and train a predictive model on the resulting table (Kanter and Veeramachaneni 2015). This approach, however, not only comes with potentially high upfront costs of joining potentially many large tables but also the structure in a relational database is ignored, since all data is represented as a single flat table in the model. This actually feels like the only valid path. It also gives you an abstraction layer such that you can modify your schema without having to retrain your models. Spending 20 minutes creating a view upfront could save you a lifetime of getting a NN to sort this out for you. I understand the arguments where preprocessing is pointless (i.e. cleaning up images for machine vision), but I don't think those hold in this domain where we are dealing with an arbitrary number of dimensions.