A rule of thumb for performing propositionalization is starting bottom-up by under-aggregating

The extreme example of starting bottom-up would be to simply transpose the dataset for the rows we aggregate towards, but while that retains all data points with the information they carry, you’ll find that there are some common issues with that approach. An example could be that on the chosen aggregate level (ex. order_id), there are varying number of rows (ex. line_item_id’s). This alone often forces us to perform some level of aggregation, so that the schema is fixed and not jagged.

The broad methodology would be something like: “coarsen only when a concrete obstacle forces it”. We iterate this process until the schema satisfies our constraints. This allows us to aggregate to the least-possible level, therefore retaining as much granularity as we possibly can.