{ "id": 501938, "title": "Drift phenomenon in the distribution of training and test dataset", "url": "/competitions/leap-atmospheric-physics-ai-climsim/discussion/501938", "author_name": "Zhuoqun Li", "post_date": "2024-05-11T12:10:17.218000", "votes": 4, "comment_count": 1, "views": 0, "content": "

I noticed significant differences in the probability density distribution of certain features between the training and test sets, such as the 'state_q0001_0' feature below. How to handle these features? Should they be deleted, retained, or subjected to some transformation?
\n

", "messages": [ { "id": 2807515, "postDate": "2024-05-11T17:46:16.477Z", "content": "

'should they'- no one can answer this before experimenting and finding out, which is what I suggest you do (and I will do myself probably at some point)
\nHowever, as long as the test values are included in the train distribution (which seems to be the case here), I expect the models to generalize well. The problems usually start to arise when the models need to generalize to never-seen-before values.

", "rawMarkdown": "'should they'- no one can answer this before experimenting and finding out, which is what I suggest you do (and I will do myself probably at some point)\nHowever, as long as the test values are included in the train distribution (which seems to be the case here), I expect the models to generalize well. The problems usually start to arise when the models need to generalize to never-seen-before values.", "votes": 3 }, { "id": 2806942, "postDate": "2024-05-11T12:10:17.220Z", "content": "

I noticed significant differences in the probability density distribution of certain features between the training and test sets, such as the 'state_q0001_0' feature below. How to handle these features? Should they be deleted, retained, or subjected to some transformation?
\n

", "rawMarkdown": "I noticed significant differences in the probability density distribution of certain features between the training and test sets, such as the 'state_q0001_0' feature below. How to handle these features? Should they be deleted, retained, or subjected to some transformation?\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11486426%2F0bff91430df41cf60a809249b8911bd0%2Fstate_q0001_0.png?generation=1715429399685794&alt=media)", "votes": 3 } ], "comments": [ { "id": 2807515, "author_name": "greySnow", "author_url": "", "post_date": "2024-05-11T17:46:16.477000", "content": "

'should they'- no one can answer this before experimenting and finding out, which is what I suggest you do (and I will do myself probably at some point)
\nHowever, as long as the test values are included in the train distribution (which seems to be the case here), I expect the models to generalize well. The problems usually start to arise when the models need to generalize to never-seen-before values.

", "votes": 3, "replies": [] } ], "raw_markdown_by_id": { "2807515": "'should they'- no one can answer this before experimenting and finding out, which is what I suggest you do (and I will do myself probably at some point)\nHowever, as long as the test values are included in the train distribution (which seems to be the case here), I expect the models to generalize well. The problems usually start to arise when the models need to generalize to never-seen-before values.", "2806942": "I noticed significant differences in the probability density distribution of certain features between the training and test sets, such as the 'state_q0001_0' feature below. How to handle these features? Should they be deleted, retained, or subjected to some transformation?\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11486426%2F0bff91430df41cf60a809249b8911bd0%2Fstate_q0001_0.png?generation=1715429399685794&alt=media)" } }