{
  "id": 76984,
  "title": "CV vs LB",
  "url": "/competitions/vsb-power-line-fault-detection/discussion/76984",
  "author_name": "",
  "post_date": "2019-01-08T13:56:21.864667800Z",
  "votes": 8,
  "comment_count": 7,
  "views": 0,
  "content": "<p>5 Fold CV: 0.5504775509, LB: 0.518\n5 Fold CV: 0.5779440177, LB: 0.557\n5 Fold CV: 0.6426664258, LB: 0.526\n5 Fold CV: 0.617476346, LB: 0.577\n5 Fold CV: 0.6254808907, LB: 0.615\n5 Fold CV: 0.6137797941, LB: 0.608\n5 Fold CV: 0.6510671877, LB: 0.565</p>\n\n<p>Its hard to see any correlation. Any ideas?</p>",
  "messages": [
    {
      "id": "452286",
      "postDate": "01/08/2019 13:56:21",
      "content": "<p>5 Fold CV: 0.5504775509, LB: 0.518\n5 Fold CV: 0.5779440177, LB: 0.557\n5 Fold CV: 0.6426664258, LB: 0.526\n5 Fold CV: 0.617476346, LB: 0.577\n5 Fold CV: 0.6254808907, LB: 0.615\n5 Fold CV: 0.6137797941, LB: 0.608\n5 Fold CV: 0.6510671877, LB: 0.565</p>\n\n<p>Its hard to see any correlation. Any ideas?</p>",
      "rawMarkdown": "5 Fold CV: 0.5504775509, LB: 0.518\n5 Fold CV: 0.5779440177, LB: 0.557\n5 Fold CV: 0.6426664258, LB: 0.526\n5 Fold CV: 0.617476346, LB: 0.577\n5 Fold CV: 0.6254808907, LB: 0.615\n5 Fold CV: 0.6137797941, LB: 0.608\n5 Fold CV: 0.6510671877, LB: 0.565\n\nIts hard to see any correlation. Any ideas?",
      "votes": null
    },
    {
      "id": "452320",
      "postDate": "01/08/2019 15:08:55",
      "content": "<p>Same problem here! But my local scores are more in the region of 0.68. I'm using a stratified k-fold by stratifying on the number of defects per measurement.</p>",
      "rawMarkdown": "Same problem here! But my local scores are more in the region of 0.68. I'm using a stratified k-fold by stratifying on the number of defects per measurement.",
      "votes": null
    },
    {
      "id": "452609",
      "postDate": "01/09/2019 00:16:02",
      "content": "<p>I have 5 Fold CV: 0.76 LB: about 0.63</p>",
      "rawMarkdown": "I have 5 Fold CV: 0.76 LB: about 0.63",
      "votes": null
    },
    {
      "id": "454398",
      "postDate": "01/11/2019 15:23:49",
      "content": "<p>Similar situation . I am working with CNN. May be cut off is in different place on test set ? </p>",
      "rawMarkdown": "Similar situation . I am working with CNN. May be cut off is in different place on test set ?",
      "votes": null
    },
    {
      "id": "456184",
      "postDate": "01/15/2019 09:50:56",
      "content": "<p>For my NN model, 10 fold CV: 0.644, LB:0.606. (gap: 0.038)\nAnd also fluctuate the relation of CV and LB every time.</p>",
      "rawMarkdown": "For my NN model, 10 fold CV: 0.644, LB:0.606. (gap: 0.038)\nAnd also fluctuate the relation of CV and LB every time.",
      "votes": null
    },
    {
      "id": "458357",
      "postDate": "01/19/2019 12:43:47",
      "content": "<p>10 fold CV: 0.690, LB:0.624<br>\nmy model is based on CapsuleNet.</p>",
      "rawMarkdown": "10 fold CV: 0.690, LB:0.624<br>\nmy model is based on CapsuleNet.",
      "votes": null
    },
    {
      "id": "459818",
      "postDate": "01/22/2019 12:23:32",
      "content": "<p>With 5 fold (split by id_measurement) my gap it's between 0.06 and 0.12, depending on the selected features.</p>",
      "rawMarkdown": "With 5 fold (split by id_measurement) my gap it's between 0.06 and 0.12, depending on the selected features.",
      "votes": null
    },
    {
      "id": "469793",
      "postDate": "02/11/2019 20:13:32",
      "content": "<p>Using LSTM architecture as mentioned here : <a href=\"https://www.kaggle.com/braquino/5-fold-lstm-attention-fully-commented-0-694\">https://www.kaggle.com/braquino/5-fold-lstm-attention-fully-commented-0-694</a></p>\n\n<p>5 Fold CV : 0.7306 LB : 0.67\n5 Fold CV : 0.7 LB : 0.689\n5 Fold CV : 0.712432 LB : 0.707\n5 Fold CV : 0.729 LB : 0.643\n5 Fold CV : 0.7487 LB : 0.658\n5 Fold CV : 0.715242 LB : 0.677\n5 Fold CV : 0.7268823 LB : 0.67</p>",
      "rawMarkdown": "Using LSTM architecture as mentioned here : https://www.kaggle.com/braquino/5-fold-lstm-attention-fully-commented-0-694\n\n5 Fold CV : 0.7306 LB : 0.67\n5 Fold CV : 0.7 LB : 0.689\n5 Fold CV : 0.712432 LB : 0.707\n5 Fold CV : 0.729 LB : 0.643\n5 Fold CV : 0.7487 LB : 0.658\n5 Fold CV : 0.715242 LB : 0.677\n5 Fold CV : 0.7268823 LB : 0.67",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 452320,
      "author_name": "maxhalford",
      "author_url": "",
      "post_date": "01/08/2019 15:08:55",
      "content": "<p>Same problem here! But my local scores are more in the region of 0.68. I'm using a stratified k-fold by stratifying on the number of defects per measurement.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 452609,
      "author_name": "onodera",
      "author_url": "",
      "post_date": "01/09/2019 00:16:02",
      "content": "<p>I have 5 Fold CV: 0.76 LB: about 0.63</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 454398,
      "author_name": "jonasmatuzas",
      "author_url": "",
      "post_date": "01/11/2019 15:23:49",
      "content": "<p>Similar situation . I am working with CNN. May be cut off is in different place on test set ? </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 456184,
      "author_name": "kenmatsu4",
      "author_url": "",
      "post_date": "01/15/2019 09:50:56",
      "content": "<p>For my NN model, 10 fold CV: 0.644, LB:0.606. (gap: 0.038)\nAnd also fluctuate the relation of CV and LB every time.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 458357,
      "author_name": "phalanx",
      "author_url": "",
      "post_date": "01/19/2019 12:43:47",
      "content": "<p>10 fold CV: 0.690, LB:0.624<br>\nmy model is based on CapsuleNet.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 459818,
      "author_name": "druidh",
      "author_url": "",
      "post_date": "01/22/2019 12:23:32",
      "content": "<p>With 5 fold (split by id_measurement) my gap it's between 0.06 and 0.12, depending on the selected features.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 469793,
      "author_name": "harshit92",
      "author_url": "",
      "post_date": "02/11/2019 20:13:32",
      "content": "<p>Using LSTM architecture as mentioned here : <a href=\"https://www.kaggle.com/braquino/5-fold-lstm-attention-fully-commented-0-694\">https://www.kaggle.com/braquino/5-fold-lstm-attention-fully-commented-0-694</a></p>\n\n<p>5 Fold CV : 0.7306 LB : 0.67\n5 Fold CV : 0.7 LB : 0.689\n5 Fold CV : 0.712432 LB : 0.707\n5 Fold CV : 0.729 LB : 0.643\n5 Fold CV : 0.7487 LB : 0.658\n5 Fold CV : 0.715242 LB : 0.677\n5 Fold CV : 0.7268823 LB : 0.67</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "452286": "5 Fold CV: 0.5504775509, LB: 0.518\n5 Fold CV: 0.5779440177, LB: 0.557\n5 Fold CV: 0.6426664258, LB: 0.526\n5 Fold CV: 0.617476346, LB: 0.577\n5 Fold CV: 0.6254808907, LB: 0.615\n5 Fold CV: 0.6137797941, LB: 0.608\n5 Fold CV: 0.6510671877, LB: 0.565\n\nIts hard to see any correlation. Any ideas?",
    "452320": "Same problem here! But my local scores are more in the region of 0.68. I'm using a stratified k-fold by stratifying on the number of defects per measurement.",
    "452609": "I have 5 Fold CV: 0.76 LB: about 0.63",
    "454398": "Similar situation . I am working with CNN. May be cut off is in different place on test set ?",
    "456184": "For my NN model, 10 fold CV: 0.644, LB:0.606. (gap: 0.038)\nAnd also fluctuate the relation of CV and LB every time.",
    "458357": "10 fold CV: 0.690, LB:0.624<br>\nmy model is based on CapsuleNet.",
    "459818": "With 5 fold (split by id_measurement) my gap it's between 0.06 and 0.12, depending on the selected features.",
    "469793": "Using LSTM architecture as mentioned here : https://www.kaggle.com/braquino/5-fold-lstm-attention-fully-commented-0-694\n\n5 Fold CV : 0.7306 LB : 0.67\n5 Fold CV : 0.7 LB : 0.689\n5 Fold CV : 0.712432 LB : 0.707\n5 Fold CV : 0.729 LB : 0.643\n5 Fold CV : 0.7487 LB : 0.658\n5 Fold CV : 0.715242 LB : 0.677\n5 Fold CV : 0.7268823 LB : 0.67"
  },
  "source": "meta"
}