{
  "id": 175145,
  "title": "Tabular data augmentation",
  "url": "/competitions/osic-pulmonary-fibrosis-progression/discussion/175145",
  "author_name": "",
  "post_date": "2020-08-17T09:26:37.118020700Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi<br>\nI'm not used to working with tabular data so I'm wondering if we can perform some augmentation during the training. I've made some searches from previous competitions with tabular data and it seems that augmentation are not pertinent. However each competition is different, especially this one where there is a limited number of FVC (with high variations) for each patient.<br>\nI was thinking of adding some (gaussian?) noise to FVC at each epoch to get a stronger model. <br>\nWhat are your thoughts about that ? Does that make any sense ? </p>",
  "messages": [
    {
      "id": "973392",
      "postDate": "08/17/2020 09:26:37",
      "content": "<p>Hi<br>\nI'm not used to working with tabular data so I'm wondering if we can perform some augmentation during the training. I've made some searches from previous competitions with tabular data and it seems that augmentation are not pertinent. However each competition is different, especially this one where there is a limited number of FVC (with high variations) for each patient.<br>\nI was thinking of adding some (gaussian?) noise to FVC at each epoch to get a stronger model. <br>\nWhat are your thoughts about that ? Does that make any sense ? </p>",
      "rawMarkdown": "Hi\nI'm not used to working with tabular data so I'm wondering if we can perform some augmentation during the training. I've made some searches from previous competitions with tabular data and it seems that augmentation are not pertinent. However each competition is different, especially this one where there is a limited number of FVC (with high variations) for each patient.\nI was thinking of adding some (gaussian?) noise to FVC at each epoch to get a stronger model. \nWhat are your thoughts about that ? Does that make any sense ?",
      "votes": null
    },
    {
      "id": "987071",
      "postDate": "08/27/2020 01:07:22",
      "content": "<p>Adding gaussian noise to the FVC will affect the confidence that your model outputs, ie) the std you choose will set your prior belief of the uncertainty. However, they do tell us that there is a measurement uncertainty in the FVC, and that is why the confidence is capped at 70ml. So using that value could be useful. I had the same idea and I will definitely try it. The use of this noise also depends on how you calculate the confidence</p>",
      "rawMarkdown": "Adding gaussian noise to the FVC will affect the confidence that your model outputs, ie) the std you choose will set your prior belief of the uncertainty. However, they do tell us that there is a measurement uncertainty in the FVC, and that is why the confidence is capped at 70ml. So using that value could be useful. I had the same idea and I will definitely try it. The use of this noise also depends on how you calculate the confidence",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 987071,
      "author_name": "samklein",
      "author_url": "",
      "post_date": "08/27/2020 01:07:22",
      "content": "<p>Adding gaussian noise to the FVC will affect the confidence that your model outputs, ie) the std you choose will set your prior belief of the uncertainty. However, they do tell us that there is a measurement uncertainty in the FVC, and that is why the confidence is capped at 70ml. So using that value could be useful. I had the same idea and I will definitely try it. The use of this noise also depends on how you calculate the confidence</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "973392": "Hi\nI'm not used to working with tabular data so I'm wondering if we can perform some augmentation during the training. I've made some searches from previous competitions with tabular data and it seems that augmentation are not pertinent. However each competition is different, especially this one where there is a limited number of FVC (with high variations) for each patient.\nI was thinking of adding some (gaussian?) noise to FVC at each epoch to get a stronger model. \nWhat are your thoughts about that ? Does that make any sense ?",
    "987071": "Adding gaussian noise to the FVC will affect the confidence that your model outputs, ie) the std you choose will set your prior belief of the uncertainty. However, they do tell us that there is a measurement uncertainty in the FVC, and that is why the confidence is capped at 70ml. So using that value could be useful. I had the same idea and I will definitely try it. The use of this noise also depends on how you calculate the confidence"
  },
  "source": "meta"
}