{ "id": 100906, "title": "which CV split is better", "url": "/competitions/recursion-cellular-image-classification/discussion/100906", "author_name": "valencebond", "post_date": "2019-07-22T03:08:04.358000", "votes": 5, "comment_count": 0, "views": 0, "content": "

i try to split trainset and valid set by cell type as below，\"\"

\n\n
    \n
  1. When i use fold0 as validation, others as train, there is a big gap nearly 0.3 between valid performance and leaderboard.
    2. \n
  2. But when i random split train and valid set, there is no such gap.
    3. \n
\n\n

I am so confused. I think the split 1 is more reasonable, because this is how the training set and test set are split.

", "messages": [ { "id": 581503, "postDate": "2019-07-22T03:08:04.357Z", "content": "

i try to split trainset and valid set by cell type as below，\"\"

\n\n
    \n
  1. When i use fold0 as validation, others as train, there is a big gap nearly 0.3 between valid performance and leaderboard.
    2. \n
  2. But when i random split train and valid set, there is no such gap.
    3. \n
\n\n

I am so confused. I think the split 1 is more reasonable, because this is how the training set and test set are split.

", "rawMarkdown": "i try to split trainset and valid set by cell type as below，![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3119385%2F4b5b2c32d21742c12fd55b9296f4dd9f%2FScreen%20Shot%202019-07-22%20at%2011.00.58%20AM.png?generation=1563764483925090&alt=media)\n\n1. When i use fold0 as validation, others as train, there is a big gap nearly 0.3 between valid performance and leaderboard. \n2. But when i random split train and valid set, there is no such gap. \n\nI am so confused. I think the split 1 is more reasonable, because this is how the training set and test set are split.", "votes": 5 } ], "comments": [], "raw_markdown_by_id": { "581503": "i try to split trainset and valid set by cell type as below，![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3119385%2F4b5b2c32d21742c12fd55b9296f4dd9f%2FScreen%20Shot%202019-07-22%20at%2011.00.58%20AM.png?generation=1563764483925090&alt=media)\n\n1. When i use fold0 as validation, others as train, there is a big gap nearly 0.3 between valid performance and leaderboard. \n2. But when i random split train and valid set, there is no such gap. \n\nI am so confused. I think the split 1 is more reasonable, because this is how the training set and test set are split." } }