question related to train_test_split

[unknown007007007]

Thanks for your work and vivid description.
But I have some questions in relaiton to train_test_split.
First, I think in proc_data/train_test_split.json, you split your data according to user_id (90%,10%).
And as of knowledge/user.klg, you mention that you only use the limited data in knowledge/begin_data.json to prevent info leakage.
Then I have some questions.

1. Did you use the data of the training set's users interaction(begin_data.json) in both user_profile and training?
2. Did you use the data of the test set's users interaction(begin_data.json) in both user_profile and training?
   Or you just use it to complete user_profile and simply delete the others in both traing and testing?
3. Why do you conduct the user_profile in training set only use limited data,even although you can use more data to make it more accurate.
4. And I sincerely hope to know how you get the hint to construct training and testing data in this way. Because for me, intuitively, I will choose to split all users' data ,which means evey user has some data in both training and testing, and I may use all of the data in the training set (or a fixed interaction number ,like40) to make user_profile, and finally get the result.

Thanks very much if you can provide some suggestion with me.