What is the best way to organize the datasets for my task?

I am working on a task — geolocation estimation of Twitter users by using tweets only. I collected tweets (and users) from more than 6000 people in Twitter. Each user is associated with a city.

In the dataset, number of samples (or users) for each city depends on the city size. (i.e. If the city A is more populous than the city B, the city A has more users in the dataset.) This seems fair, but it creates an unbalanced dataset.

Right now, I am planing to collect another dataset; a dataset that is more balanced (i.e. there will be the almost the same amount of users for each city although there will still be more users in big cities). Doing this makes sense or should I continue with the unbalanced dataset? What approach would it be good for that task?

When you uniformly take samples from a society, definitely chance of selecting from cities is directly related to their population. Therefore, more users will be selected from more populous cities and it's one of the most important characters of the problem you trying to solve. I think if you want to balance the data-set, you ignore one this important character of your data and also your problem.

I strongly recommend to continue with the unbalanced data-set and handle it by choosing an appropriate loss function and evaluation method.

Disclaimer:

If you use python, PyCM module can help you to find out these metrics.

Here is a simple code to get the recommended parameters from this module:

>>> from pycm import *

>>> cm = ConfusionMatrix(matrix={"Class1": {"Class1": 1, "Class2":2}, "Class2": {"Class1": 0, "Class2": 5}})

>>> print(cm.recommended_list)
["Kappa", "SOA1(Landis & Koch)", "SOA2(Fleiss)", "SOA3(Altman)", "SOA4(Cicchetti)", "CEN", "MCEN", "MCC", "J", "Overall J", "Overall MCC", "Overall CEN", "Overall MCEN", "AUC", "AUCI", "G", "DP", "DPI", "GI"]


After that, each of these parameters you want to use as the loss function can be used as follows:

>>> y_pred = model.predict      #the prediction of the implemented model

>>> y_actu = data.target        #data labels

>>> cm = ConfusionMatrix(y_actu, y_pred)

>>> loss = cm.Kappa             #or any other parameter (Example: cm.SOA1)