{
  "id": 507948,
  "title": "Public 437/ Private 198 solution ",
  "url": "/competitions/home-credit-credit-risk-model-stability/discussion/507948",
  "author_name": "",
  "post_date": "2024-05-28T00:41:40.892172900Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Initially, I tried to get a good score with just one lgbm model due to the selection of features and hyperparameters, very quickly realized that this was not enough and moved on to ensembles. </p>\n<p>The first thing I noticed is that the greater the difference between сatboost and lgm, the more they \"help\" each other. This gave me the opportunity to get a great  score. After that, the well-known <strong>this is the way</strong> appeared😂 of which I discovered metric hack, but immediately realized that you can't get carried away with it, so as not to retrain for the public. </p>\n<p>The final boost was soon given by a voting ensemble of two lgbms with different types of boosting. In general, this whole competition is an illustrative example of the fact that you can not get carried away with hyperparameter selection)))<br>\nWrite your findings in the comments, it will be interesting to listen!!!</p>",
  "messages": [
    {
      "id": "2840074",
      "postDate": "05/28/2024 00:41:40",
      "content": "<p>Initially, I tried to get a good score with just one lgbm model due to the selection of features and hyperparameters, very quickly realized that this was not enough and moved on to ensembles. </p>\n<p>The first thing I noticed is that the greater the difference between сatboost and lgm, the more they \"help\" each other. This gave me the opportunity to get a great  score. After that, the well-known <strong>this is the way</strong> appeared😂 of which I discovered metric hack, but immediately realized that you can't get carried away with it, so as not to retrain for the public. </p>\n<p>The final boost was soon given by a voting ensemble of two lgbms with different types of boosting. In general, this whole competition is an illustrative example of the fact that you can not get carried away with hyperparameter selection)))<br>\nWrite your findings in the comments, it will be interesting to listen!!!</p>",
      "rawMarkdown": "Initially, I tried to get a good score with just one lgbm model due to the selection of features and hyperparameters, very quickly realized that this was not enough and moved on to ensembles. \n\nThe first thing I noticed is that the greater the difference between сatboost and lgm, the more they \"help\" each other. This gave me the opportunity to get a great  score. After that, the well-known **this is the way** appeared😂 of which I discovered metric hack, but immediately realized that you can't get carried away with it, so as not to retrain for the public. \n\nThe final boost was soon given by a voting ensemble of two lgbms with different types of boosting. In general, this whole competition is an illustrative example of the fact that you can not get carried away with hyperparameter selection)))\nWrite your findings in the comments, it will be interesting to listen!!!",
      "votes": null
    },
    {
      "id": "2840131",
      "postDate": "05/28/2024 01:42:17",
      "content": "<p>CatBoost seems to be the one giving the least overfitting. Here's my summary:</p>\n<table>\n<thead>\n<tr>\n<th>Model</th>\n<th>Public</th>\n<th>Private</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Cat</td>\n<td>0.579</td>\n<td>0.509</td>\n</tr>\n<tr>\n<td>LGB</td>\n<td>0.580</td>\n<td>0.500</td>\n</tr>\n<tr>\n<td>XGB</td>\n<td>0.586</td>\n<td>0.509</td>\n</tr>\n<tr>\n<td>sklearn histgradientboost</td>\n<td>0.585</td>\n<td>0.501</td>\n</tr>\n</tbody>\n</table>\n<p>Even when combining the results. Combining with catboost gave better results, XGB and sklearn worsened the results somehow.</p>",
      "rawMarkdown": "CatBoost seems to be the one giving the least overfitting. Here's my summary:\n| Model | Public | Private |\n| --- | --- | --- |\n| Cat | 0.579| 0.509 |\n| LGB | 0.580 | 0.500 |\n| XGB |0.586 | 0.509 |\n| sklearn histgradientboost | 0.585 | 0.501 |\n\nEven when combining the results. Combining with catboost gave better results, XGB and sklearn worsened the results somehow.",
      "votes": null
    },
    {
      "id": "2840242",
      "postDate": "05/28/2024 03:32:53",
      "content": "<p>Thx for sharing </p>",
      "rawMarkdown": "Thx for sharing",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2840131,
      "author_name": "soonjynnchu",
      "author_url": "",
      "post_date": "05/28/2024 01:42:17",
      "content": "<p>CatBoost seems to be the one giving the least overfitting. Here's my summary:</p>\n<table>\n<thead>\n<tr>\n<th>Model</th>\n<th>Public</th>\n<th>Private</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Cat</td>\n<td>0.579</td>\n<td>0.509</td>\n</tr>\n<tr>\n<td>LGB</td>\n<td>0.580</td>\n<td>0.500</td>\n</tr>\n<tr>\n<td>XGB</td>\n<td>0.586</td>\n<td>0.509</td>\n</tr>\n<tr>\n<td>sklearn histgradientboost</td>\n<td>0.585</td>\n<td>0.501</td>\n</tr>\n</tbody>\n</table>\n<p>Even when combining the results. Combining with catboost gave better results, XGB and sklearn worsened the results somehow.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2840242,
          "author_name": "oxygen22831",
          "author_url": "",
          "post_date": "05/28/2024 03:32:53",
          "content": "<p>Thx for sharing </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2840074": "Initially, I tried to get a good score with just one lgbm model due to the selection of features and hyperparameters, very quickly realized that this was not enough and moved on to ensembles. \n\nThe first thing I noticed is that the greater the difference between сatboost and lgm, the more they \"help\" each other. This gave me the opportunity to get a great  score. After that, the well-known **this is the way** appeared😂 of which I discovered metric hack, but immediately realized that you can't get carried away with it, so as not to retrain for the public. \n\nThe final boost was soon given by a voting ensemble of two lgbms with different types of boosting. In general, this whole competition is an illustrative example of the fact that you can not get carried away with hyperparameter selection)))\nWrite your findings in the comments, it will be interesting to listen!!!",
    "2840131": "CatBoost seems to be the one giving the least overfitting. Here's my summary:\n| Model | Public | Private |\n| --- | --- | --- |\n| Cat | 0.579| 0.509 |\n| LGB | 0.580 | 0.500 |\n| XGB |0.586 | 0.509 |\n| sklearn histgradientboost | 0.585 | 0.501 |\n\nEven when combining the results. Combining with catboost gave better results, XGB and sklearn worsened the results somehow.",
    "2840242": "Thx for sharing"
  },
  "source": "meta"
}