Hello all, just finished Part 1 for fast ai. As a project, I have a dataset that is at order/item level and I am trying to predict a binary output at the order level. I am mainly confused about how to set this problem up. I can aggregate values at the order level and then predict the output, but I feel like I am losing valuable insight by not letting the model know which items were part of the order. I can convert the item-level data to columns and add 1 if the item was part of the order 0 if not (similar to one hot encoding but with multiple items), but that would add close to 30K columns to my tabular data.
My initial intuition says this could potentially be solved with an NLP + tabular data approach but I am not entirely sure how to set it up. Would you concatenate all the items into a single column, and then train an NLP model[Where the sentence is the order, and the words are the items] with the order level binary output, then take those embedding and include them in the tabular data to solve the final problem. If so, how would the embedding come into the new model architecture, if not what approach do you recommend? Any insight would be greatly appreciated.
Example Dataset: x = some continuous/categorical variable
Order# | Item# | col 1 | col2 | col3 | target |
A 1 x x x 1
A 2 x x x 1
A 3 x x x 1
A 4 x x x 1
B 2 x x x 0
B 3 x x x 0
C 1 x x x 1
C 4 x x x 1