Hello, I’ve ran into the issue with tabular data. All of the examples listed have some sort of one-item per row. As for me, i have aggregated data like in the example:
item count | average price | item name
1 | 5 | bag
4 | 2 | box
2 | 3 | bag
Obviously, we should somehow increase the importance of the rows so that row 3 has to be twice more important than row 1. However i didn’t find any info on how to increase the learning rate for those accordingly and I dont want to just break that info into several different rows without adding any extra value. any ideas on how to solve this?
in this example im trying to predict the price for the Item Name. in this case, “Bag” has two enties where the second one has two items with average price of 3, so it has to be twice as important compared to the first one.
yes, thats correct. Just in case: Im simplifying the real task as in reality im working on a hybrid nn working both with text and tabular data, but this is so far the only complication I ran into.
I’m prefacing this with I’m sure I’m being slow here. Why would you need to add weight to the line item? What information will you have at time of inference? If it’s the item name and a quantity, the network should be able to handle those to come up with the average price? Have you given this a try with no feature engineering/massaging?
Again, I’m sure I’m missing something blatantly obvious, so my apologies in advance
