Multivariate time-series forecasting with LSTM

Question

I have dataset (near 4 million observations) with 4 features:

• date
• id_shop (table of shops consist only names of 61 shops)
• id_item (there are names of 22 000 items)
• item_cnt_day - number of products sold

item_cnt_day is my target. I have to predict monthly products sold in each shop and for each item.

I want to try LSTM and have questions:

1. How to encode id_shop and id_item for LSTM? One-hot encoding will give me 22000 + 61 extra features. Аre there any alternatives?
2. Is it good idea use LSTM for this problem?
3. What input is better to use for lstm?

Answer 1

1. Use Distributed Representation. Create a random vector of around 50 (take this number with a pinch of salt, get to the decent number while training the network) dimensions for each item and shop and learn them as parameters of the neural network

2. Maybe - If you have good amount of computation power, you can try. No - if not. I don't see why a carefully designed MLP wont be able to do the job but as is the answer with many deep learning questions - Try and see. But start with MLP. This isnt a time series data unless you are trying to model it based upon the sales of previous days as well in which case (and which should be the case. i think the sales would indeed depend upon previous sales) use LSTM (maybe Bi-LSTM)

3. You dont have many inputs. Use them all. if you were asking what kind of problems are LSTMs used for - they are used for time series data that would change sense with change in the order they are presented in