Models don't perform as expected when trained on a large train set

[atkm]

From Kaggle submissions (private scores):

Model 3:

• Half of train.csv: 0.4241803
• train_small.csv: 0.4123864

Model 1:

• Half of train.csv: 0.4172254
• train_small.csv: 0.4168748

Run an experiment to pick an optimal training set size. One way to reduce the training set is to remove rows with (device_id, site/app_id) pairs that do not appear on the test set.

[atkm]

Maybe the model needs a different regularization parameter when fitting to a large dataset?

[atkm]

A simple classifier like logistic regression shouldn't require as much data as a more complex one like decision trees.