How do I prepare a multi group time series dataset into a supervised learning one?

Question

ML newbie here, I have a time series dataset that looks like this:

   ID   Hour    P1      P2      P3      P4     Target
    1   1       95.0    36.11   75.33   19.0    0
    1   2       95.0    36.11   75.33   19.0    0
    1   3       99.0    36.11   86.00   22.0    1

It has the data of multiple users grouped by their IDs and the task is to predict the target label. I'm planning to convert it into a supervised learning task by appending shifted columns and use an LSTM to predict the target value for the next hour.

My questions are:

1. Since there are multiple users, would the prediction be impacted if no separation between users is made? And if so, how do I go about it? Do I just add rows of zeros between users equal to the number of shifts?
2. Is using a time series approach recommended for this task or would a regression approach be more suited?

Answer 1

For your first question, more details about your task would be helpful to effectively answer it, but, from what I can gather, I think the best way is for your input to the LSTM to be all of the features for all of the users of a given time step concatenated together with no zero. Again, you didn't specify the specific aspect of your problem so that really the best I can give you.

As for your second question, I think you should do more analysis on your data before you come to the conclusion that you need a big, slow, and expensive neural network to solve your problem. Things you might want to do

1. Try PCA and project your data into a space of two principal components to visualize its structure
2. Try machine learning techniques like XGBoost, Kernel Machines, shallow neural networks, or clustering

Generally, one's choice of a specific neural network should not be arbitrary. It should be deliberately chosen, based on hard evidence, to suit the structure of the data and its complexity.