{
  "id": 343949,
  "title": "Any advices to improve the LB to the 0.799??",
  "url": "/competitions/amex-default-prediction/discussion/343949",
  "author_name": "",
  "post_date": "2022-08-13T08:03:14.603103500Z",
  "votes": 1,
  "comment_count": 6,
  "views": 0,
  "content": "<p>hey guys, hope u all enjoy this game.</p>\n<p>By the way, my limit seems in the 0.799 of LB.</p>\n<p>what now i have is a model about 900 features and a dart LGBM seem can reach 0.79800000000…. ,<br>\na cat boost model 0.795 LB(Seems cat boost not attractive in thie compe)<br>\ntry to blending 3 different seeds of  one same dart model, and not really useful…</p>\n<p>And i didnt do the stacking,  after that, possivle can get 0.7983- 0.7985</p>\n<p>now i kind of stucking, may be some blind spot on me ,you guys can see</p>",
  "messages": [
    {
      "id": "1896847",
      "postDate": "08/13/2022 08:03:14",
      "content": "<p>hey guys, hope u all enjoy this game.</p>\n<p>By the way, my limit seems in the 0.799 of LB.</p>\n<p>what now i have is a model about 900 features and a dart LGBM seem can reach 0.79800000000…. ,<br>\na cat boost model 0.795 LB(Seems cat boost not attractive in thie compe)<br>\ntry to blending 3 different seeds of  one same dart model, and not really useful…</p>\n<p>And i didnt do the stacking,  after that, possivle can get 0.7983- 0.7985</p>\n<p>now i kind of stucking, may be some blind spot on me ,you guys can see</p>",
      "rawMarkdown": "hey guys, hope u all enjoy this game.\n\nBy the way, my limit seems in the 0.799 of LB.\n\nwhat now i have is a model about 900 features and a dart LGBM seem can reach 0.79800000000.... ,\na cat boost model 0.795 LB(Seems cat boost not attractive in thie compe)\ntry to blending 3 different seeds of  one same dart model, and not really useful...\n\nAnd i didnt do the stacking,  after that, possivle can get 0.7983- 0.7985\n\n \nnow i kind of stucking, may be some blind spot on me ,you guys can see",
      "votes": null
    },
    {
      "id": "1896972",
      "postDate": "08/13/2022 10:32:07",
      "content": "<p><a href=\"https://www.kaggle.com/competitions/amex-default-prediction/discussion/338906\" target=\"_blank\">LB 0.798 is possible with catboost</a>, but my advice would be to add a neural network model to the ensemble. </p>",
      "rawMarkdown": "[LB 0.798 is possible with catboost](https://www.kaggle.com/competitions/amex-default-prediction/discussion/338906), but my advice would be to add a neural network model to the ensemble.",
      "votes": null
    },
    {
      "id": "1897027",
      "postDate": "08/13/2022 11:48:25",
      "content": "<p>can i have more details please? i use optuna on my catboost, but never have one that cv can over 0.795</p>",
      "rawMarkdown": "can i have more details please? i use optuna on my catboost, but never have one that cv can over 0.795",
      "votes": null
    },
    {
      "id": "1906454",
      "postDate": "08/19/2022 22:39:25",
      "content": "<p>Ensememble uncorrelated (as much as you can) models</p>\n<p>The Devastator.</p>",
      "rawMarkdown": "Ensememble uncorrelated (as much as you can) models\n\nThe Devastator.",
      "votes": null
    },
    {
      "id": "1906769",
      "postDate": "08/20/2022 07:25:58",
      "content": "<p>oh~ it's make sense to me, thansk a lot  <a href=\"https://www.kaggle.com/thedevastator\" target=\"_blank\">@thedevastator</a> </p>",
      "rawMarkdown": "oh~ it's make sense to me, thansk a lot  @thedevastator",
      "votes": null
    },
    {
      "id": "1907242",
      "postDate": "08/20/2022 15:48:18",
      "content": "<p>Adversarial Validation combined with shap importance worked well for me. Basically you treat the AUC and shap importance thresholds as hyperparameters then you drop the unstable features but keep the high shap even if unstable.</p>",
      "rawMarkdown": "Adversarial Validation combined with shap importance worked well for me. Basically you treat the AUC and shap importance thresholds as hyperparameters then you drop the unstable features but keep the high shap even if unstable.",
      "votes": null
    },
    {
      "id": "1907243",
      "postDate": "08/20/2022 15:49:24",
      "content": "<p>yes, also adding NNs helped me, because the algorithm is totally different from tree based models like catboost.</p>",
      "rawMarkdown": "yes, also adding NNs helped me, because the algorithm is totally different from tree based models like catboost.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1896972,
      "author_name": "ambrosm",
      "author_url": "",
      "post_date": "08/13/2022 10:32:07",
      "content": "<p><a href=\"https://www.kaggle.com/competitions/amex-default-prediction/discussion/338906\" target=\"_blank\">LB 0.798 is possible with catboost</a>, but my advice would be to add a neural network model to the ensemble. </p>",
      "votes": null,
      "replies": [
        {
          "id": 1897027,
          "author_name": "taylor1224",
          "author_url": "",
          "post_date": "08/13/2022 11:48:25",
          "content": "<p>can i have more details please? i use optuna on my catboost, but never have one that cv can over 0.795</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1907243,
          "author_name": "pabuoro",
          "author_url": "",
          "post_date": "08/20/2022 15:49:24",
          "content": "<p>yes, also adding NNs helped me, because the algorithm is totally different from tree based models like catboost.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1906454,
      "author_name": "thedevastator",
      "author_url": "",
      "post_date": "08/19/2022 22:39:25",
      "content": "<p>Ensememble uncorrelated (as much as you can) models</p>\n<p>The Devastator.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1906769,
          "author_name": "taylor1224",
          "author_url": "",
          "post_date": "08/20/2022 07:25:58",
          "content": "<p>oh~ it's make sense to me, thansk a lot  <a href=\"https://www.kaggle.com/thedevastator\" target=\"_blank\">@thedevastator</a> </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1907242,
      "author_name": "pabuoro",
      "author_url": "",
      "post_date": "08/20/2022 15:48:18",
      "content": "<p>Adversarial Validation combined with shap importance worked well for me. Basically you treat the AUC and shap importance thresholds as hyperparameters then you drop the unstable features but keep the high shap even if unstable.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1896847": "hey guys, hope u all enjoy this game.\n\nBy the way, my limit seems in the 0.799 of LB.\n\nwhat now i have is a model about 900 features and a dart LGBM seem can reach 0.79800000000.... ,\na cat boost model 0.795 LB(Seems cat boost not attractive in thie compe)\ntry to blending 3 different seeds of  one same dart model, and not really useful...\n\nAnd i didnt do the stacking,  after that, possivle can get 0.7983- 0.7985\n\n \nnow i kind of stucking, may be some blind spot on me ,you guys can see",
    "1896972": "[LB 0.798 is possible with catboost](https://www.kaggle.com/competitions/amex-default-prediction/discussion/338906), but my advice would be to add a neural network model to the ensemble.",
    "1897027": "can i have more details please? i use optuna on my catboost, but never have one that cv can over 0.795",
    "1906454": "Ensememble uncorrelated (as much as you can) models\n\nThe Devastator.",
    "1906769": "oh~ it's make sense to me, thansk a lot  @thedevastator",
    "1907242": "Adversarial Validation combined with shap importance worked well for me. Basically you treat the AUC and shap importance thresholds as hyperparameters then you drop the unstable features but keep the high shap even if unstable.",
    "1907243": "yes, also adding NNs helped me, because the algorithm is totally different from tree based models like catboost."
  },
  "source": "meta"
}