{
  "id": 12456,
  "title": "Big difference between local MCL score and LB one",
  "url": "/competitions/datasciencebowl/discussion/12456",
  "author_name": "",
  "post_date": "2015-02-09T20:16:35.473Z",
  "votes": 1,
  "comment_count": 2,
  "views": 1078,
  "content": "<p>Hi All,</p>\n<p>I got my 2.12 using SVC on filtered resized images and a RandomizedPCA for features selection.</p>\n<p>What puzzled me is that often the values I found locally for MCL (for which I used the tutorial code) is very different from the one in the LB.</p>\n<p>Locally the range of MCL is&nbsp; in the interval 2-4, but sometimes when I submit results I found also values from 5 to 24!!</p>\n<p>I used a modified version of tutorial code, but I prepared the train and the test data before the classification using skimage algo.</p>\n<p>thanks</p>",
  "messages": [
    {
      "id": "63825",
      "postDate": "02/09/2015 20:16:35",
      "content": "<p>Hi All,</p>\n<p>I got my 2.12 using SVC on filtered resized images and a RandomizedPCA for features selection.</p>\n<p>What puzzled me is that often the values I found locally for MCL (for which I used the tutorial code) is very different from the one in the LB.</p>\n<p>Locally the range of MCL is&nbsp; in the interval 2-4, but sometimes when I submit results I found also values from 5 to 24!!</p>\n<p>I used a modified version of tutorial code, but I prepared the train and the test data before the classification using skimage algo.</p>\n<p>thanks</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63892",
      "postDate": "02/10/2015 08:12:15",
      "content": "<p>Before I made the first submission, I was&nbsp;worried about whether the distributions of my local training and validation sets are the same with the distribution of the LB dataset, whether the conclusions found on my local dataset still hold on the LB dataset.&nbsp;</p>\n\n<p>After several times of submissions, I found the scores on my local validation sets are always consistent with the the LB scores, which of course clear my worries. For detailed information, I used 90% for training and 10% for validation. The validation set is&nbsp;not used for refining my models. They are only used for model selection and combination.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63986",
      "postDate": "02/11/2015 00:39:27",
      "content": "<p>If you get as large as 24, you need to check your submission file for issues like incorrect class labeling such as two or more class labels being incorrectly swapped.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 63892,
      "author_name": "xudongcao",
      "author_url": "",
      "post_date": "02/10/2015 08:12:15",
      "content": "<p>Before I made the first submission, I was&nbsp;worried about whether the distributions of my local training and validation sets are the same with the distribution of the LB dataset, whether the conclusions found on my local dataset still hold on the LB dataset.&nbsp;</p>\n\n<p>After several times of submissions, I found the scores on my local validation sets are always consistent with the the LB scores, which of course clear my worries. For detailed information, I used 90% for training and 10% for validation. The validation set is&nbsp;not used for refining my models. They are only used for model selection and combination.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63986,
      "author_name": "deepcnn",
      "author_url": "",
      "post_date": "02/11/2015 00:39:27",
      "content": "<p>If you get as large as 24, you need to check your submission file for issues like incorrect class labeling such as two or more class labels being incorrectly swapped.&nbsp;</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "63825": "",
    "63892": "",
    "63986": ""
  },
  "source": "meta"
}