session_id and level_group:\n
![\"For](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F690886%2Fba162ce133d1ef44c84244c7e1a6d118%2FScreenshot%202023-06-29%20at%208.51.49%20AM.png?generation=1687999968380750&alt=media\")
TLDR:
\nIn test/sample_submission.csv, there is a unseen column session_level, if you have tried to download the given jo_wilder Python api and run it on your machine, you will see it cannot be run because of missing this column.
It is not hard to show, this columns is actually the encoding of 2 columns session_id and level_group:
\n
Another thing we can confirm is that the test loop api only yield 3 iterations for each session id, because of this encoding method. This is already well known idea in this discussion, but by understanding this encoding method, we can see that it is not possible for a next iter having lower level given the same session_id.
\nIn otherwords even in the training data we are seeing some \"restarted game\": Lv1 Lv2 Lv3 Lv1' Lv2' …, but in the test loop api, they will be presented to use (Lv1 Lv1') (Lv2 Lv2') Lv3
", "messages": [ { "id": "2321877", "postDate": "06/29/2023 01:02:45", "content": "TLDR:
\nIn test/sample_submission.csv, there is a unseen column session_level, if you have tried to download the given jo_wilder Python api and run it on your machine, you will see it cannot be run because of missing this column.
It is not hard to show, this columns is actually the encoding of 2 columns session_id and level_group:
\n
Another thing we can confirm is that the test loop api only yield 3 iterations for each session id, because of this encoding method. This is already well known idea in this discussion, but by understanding this encoding method, we can see that it is not possible for a next iter having lower level given the same session_id.
\nIn otherwords even in the training data we are seeing some \"restarted game\": Lv1 Lv2 Lv3 Lv1' Lv2' …, but in the test loop api, they will be presented to use (Lv1 Lv1') (Lv2 Lv2') Lv3
", "rawMarkdown": "TLDR:\n- This is not directly related to our solution, we don't gain much advantage from this finding.\n- Create \"session_level\" column in the train.csv: https://www.kaggle.com/code/kingychiu/session-level-encoding-for-train-dataset\n\nIn test/sample_submission.csv, there is a unseen column `session_level`, if you have tried to download the given `jo_wilder` Python api and run it on your machine, you will see it cannot be run because of missing this column.\n\nIt is not hard to show, this columns is actually the encoding of 2 columns `session_id` and `level_group`:\n\n\nAnother thing we can confirm is that the test loop api only yield 3 iterations for each session id, because of this encoding method. This is already well known idea in this discussion, but by understanding this encoding method, we can see that it is not possible for a next iter having lower level given the same session_id. \n\nIn otherwords even in the training data we are seeing some \"restarted game\": Lv1 Lv2 Lv3 Lv1' Lv2' ..., but in the test loop api, they will be presented to use (Lv1 Lv1') (Lv2 Lv2') Lv3", "votes": null }, { "id": "2322062", "postDate": "06/29/2023 04:14:55", "content": "Interesting. I didn't pay much attention to that session_level. This implies that if an user goes Lv 1 > 2 > 3 >1 > 2…, you can use (Lv1 Lv1) to gain more information about subsequently groups?
You said your solution didn't benefit much from this information - does it mean you did not use this idea or that it did not improve your model?
", "rawMarkdown": "Interesting. I didn't pay much attention to that `session_level`. This implies that if an user goes Lv 1 > 2 > 3 >1 > 2..., you can use (Lv1 Lv1) to gain more information about subsequently groups?\n\nYou said your solution didn't benefit much from this information - does it mean you did not use this idea or that it did not improve your model?", "votes": null }, { "id": "2322262", "postDate": "06/29/2023 07:14:58", "content": "I think most people have this information implicitly already. Most public notebooks first filter the training data by a question and train on each question. If you want to contact the features for Q1 for Q2 Q3 etc… without sorting after the concat, you will end up having (Lv1 Lv1') (Lv2 Lv2') Lv3 in your training data as well. Therefore we might implicitly use this idea, but we are not abusing this fact as a leak.
In fact, in some of our models we sort after concat different qs/lvs, which means we enforce both the train and test data has no reverse time: (Lv1 Lv1') (Lv2 Lv2') Lv3 -> Lv1 Lv2 Lv3 Lv1 Lv2… We will have a solution post later for details.
Interesting. I didn't pay much attention to that session_level. This implies that if an user goes Lv 1 > 2 > 3 >1 > 2…, you can use (Lv1 Lv1) to gain more information about subsequently groups?
You said your solution didn't benefit much from this information - does it mean you did not use this idea or that it did not improve your model?
", "votes": null, "replies": [ { "id": 2322262, "author_name": "kingychiu", "author_url": "", "post_date": "06/29/2023 07:14:58", "content": "I think most people have this information implicitly already. Most public notebooks first filter the training data by a question and train on each question. If you want to contact the features for Q1 for Q2 Q3 etc… without sorting after the concat, you will end up having (Lv1 Lv1') (Lv2 Lv2') Lv3 in your training data as well. Therefore we might implicitly use this idea, but we are not abusing this fact as a leak.
In fact, in some of our models we sort after concat different qs/lvs, which means we enforce both the train and test data has no reverse time: (Lv1 Lv1') (Lv2 Lv2') Lv3 -> Lv1 Lv2 Lv3 Lv1 Lv2… We will have a solution post later for details.