{
  "id": 168929,
  "title": "Model Ensembles and Confidence",
  "url": "/competitions/osic-pulmonary-fibrosis-progression/discussion/168929",
  "author_name": "",
  "post_date": "2020-07-22T11:38:01.568242500Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Many (all?) of the top scoring models at the moment are using the quantile loss. Many (probably all) of us still have some way to go with extracting information from the scans. However, given that lots of kaggle competitions are won using ensembles of models, what do we think the effect will be of combining models in terms of confidence and how might we achieve that? </p>\n\n<p>In particular I'd be interested in anyone's experience with the handful of other uncertainty based competitions in this regard. </p>",
  "messages": [
    {
      "id": "939691",
      "postDate": "07/22/2020 11:38:01",
      "content": "<p>Many (all?) of the top scoring models at the moment are using the quantile loss. Many (probably all) of us still have some way to go with extracting information from the scans. However, given that lots of kaggle competitions are won using ensembles of models, what do we think the effect will be of combining models in terms of confidence and how might we achieve that? </p>\n\n<p>In particular I'd be interested in anyone's experience with the handful of other uncertainty based competitions in this regard. </p>",
      "rawMarkdown": "Many (all?) of the top scoring models at the moment are using the quantile loss. Many (probably all) of us still have some way to go with extracting information from the scans. However, given that lots of kaggle competitions are won using ensembles of models, what do we think the effect will be of combining models in terms of confidence and how might we achieve that? \n\nIn particular I'd be interested in anyone's experience with the handful of other uncertainty based competitions in this regard.",
      "votes": null
    },
    {
      "id": "948970",
      "postDate": "07/28/2020 10:54:01",
      "content": "<p>This is a very good idea, but in my experiments, the effects of various tree models like lgb and xgb are very poor. I think in the final ranking, NN methods will definitely rank high.</p>",
      "rawMarkdown": "This is a very good idea, but in my experiments, the effects of various tree models like lgb and xgb are very poor. I think in the final ranking, NN methods will definitely rank high.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 948970,
      "author_name": "xiaowangiiiii",
      "author_url": "",
      "post_date": "07/28/2020 10:54:01",
      "content": "<p>This is a very good idea, but in my experiments, the effects of various tree models like lgb and xgb are very poor. I think in the final ranking, NN methods will definitely rank high.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "939691": "Many (all?) of the top scoring models at the moment are using the quantile loss. Many (probably all) of us still have some way to go with extracting information from the scans. However, given that lots of kaggle competitions are won using ensembles of models, what do we think the effect will be of combining models in terms of confidence and how might we achieve that? \n\nIn particular I'd be interested in anyone's experience with the handful of other uncertainty based competitions in this regard.",
    "948970": "This is a very good idea, but in my experiments, the effects of various tree models like lgb and xgb are very poor. I think in the final ranking, NN methods will definitely rank high."
  },
  "source": "meta"
}