{
  "id": 77849,
  "title": "Stacking and Blending on LB",
  "url": "/competitions/elo-merchant-category-recommendation/discussion/77849",
  "author_name": "Manoj",
  "post_date": "2019-01-17T04:55:30.975000",
  "votes": 5,
  "comment_count": 10,
  "views": 0,
  "content": "<p>How much extra are you able to achieve on the LB using stacking and blending over your best single model ?</p>\n\n<p>Edit 1: \nin my case  LB 3.689 (best single model) ---&gt;&gt;  3.686 </p>\n\n<p>Edit 2:\n LB 3.685 (best single model) ---&gt;&gt;  3.674 (done by teammate @AmirH)</p>",
  "messages": [
    {
      "id": 457213,
      "postDate": "2019-01-17T04:55:30.977Z",
      "content": "<p>How much extra are you able to achieve on the LB using stacking and blending over your best single model ?</p>\n\n<p>Edit 1: \nin my case  LB 3.689 (best single model) ---&gt;&gt;  3.686 </p>\n\n<p>Edit 2:\n LB 3.685 (best single model) ---&gt;&gt;  3.674 (done by teammate @AmirH)</p>",
      "rawMarkdown": "How much extra are you able to achieve on the LB using stacking and blending over your best single model ?\n\nEdit 1: \nin my case  LB 3.689 (best single model) ---&gt;&gt;  3.686 \n\nEdit 2:\n LB 3.685 (best single model) ---&gt;&gt;  3.674 (done by teammate @AmirH)",
      "votes": 5
    },
    {
      "id": 459815,
      "postDate": "2019-01-22T12:20:42.643Z",
      "content": "<p>I haven't check my best LB among the models in my stacking, but I estimate it brings about 0.005 improvement on LB.\nIn terms of CV, stacking bring about 0.012, but part of that is very likely to be overfitting. </p>\n\n<p>I think the improvement of stacking decreases as your score moves toward lower RMSE - I remember in my very first stacking test run with LGB model at LB3.695, I could achieve stacking score LB 3.685</p>",
      "rawMarkdown": "I haven't check my best LB among the models in my stacking, but I estimate it brings about 0.005 improvement on LB.\nIn terms of CV, stacking bring about 0.012, but part of that is very likely to be overfitting. \n\nI think the improvement of stacking decreases as your score moves toward lower RMSE - I remember in my very first stacking test run with LGB model at LB3.695, I could achieve stacking score LB 3.685\n",
      "votes": 4
    },
    {
      "id": 457372,
      "postDate": "2019-01-17T09:45:43.340Z",
      "content": "<p>Very basic stacking got me 0.011 improvement.\nI will use my entire stacking and blending pipeline once i finish some more FE. \nMy single model is not as good as those i see here </p>",
      "rawMarkdown": "Very basic stacking got me 0.011 improvement.\nI will use my entire stacking and blending pipeline once i finish some more FE. \nMy single model is not as good as those i see here ",
      "votes": 1,
      "replies": [
        {
          "id": 457389,
          "postDate": "2019-01-17T10:19:00.193Z",
          "content": "<p>thanks, just wanted to have an idea as i am not able to improve my single model\ngood to know that there is a decent scope in stacking :)</p>",
          "rawMarkdown": "thanks, just wanted to have an idea as i am not able to improve my single model\ngood to know that there is a decent scope in stacking :)"
        },
        {
          "id": 459787,
          "postDate": "2019-01-22T11:19:58.107Z",
          "content": "<p>0.011 imp is very a very big imp for this competition.</p>",
          "rawMarkdown": "0.011 imp is very a very big imp for this competition."
        },
        {
          "id": 459820,
          "postDate": "2019-01-22T12:31:11.207Z",
          "content": "<p>Thanks Wang,\nI am now working on my entire pipeline, in order to run an entire solution incl. blending and stacking, and then with new LB standing - look for a teammate with strong single model.\nAlso - my CV-LB relationship remains relatively consistent with stacking.</p>",
          "rawMarkdown": "Thanks Wang,\nI am now working on my entire pipeline, in order to run an entire solution incl. blending and stacking, and then with new LB standing - look for a teammate with strong single model.\nAlso - my CV-LB relationship remains relatively consistent with stacking."
        },
        {
          "id": 463951,
          "postDate": "2019-01-31T00:13:17.267Z",
          "content": "<p>my problem is that my XGB model is doing so bad that it worsen my blending</p>",
          "rawMarkdown": "my problem is that my XGB model is doing so bad that it worsen my blending"
        }
      ]
    },
    {
      "id": 457230,
      "postDate": "2019-01-17T05:38:24.180Z",
      "content": "<p>No idea, nothing successful for me.</p>",
      "rawMarkdown": "No idea, nothing successful for me.",
      "votes": 1,
      "replies": [
        {
          "id": 457280,
          "postDate": "2019-01-17T07:22:10.533Z",
          "content": "<p>current LB is your single model , wow !</p>",
          "rawMarkdown": "current LB is your single model , wow !"
        },
        {
          "id": 457353,
          "postDate": "2019-01-17T09:10:01.750Z",
          "content": "<p>No, we use post-process trick, but we definitely won't use it for final submissions.</p>",
          "rawMarkdown": "No, we use post-process trick, but we definitely won't use it for final submissions.",
          "votes": 1
        }
      ]
    },
    {
      "id": 463124,
      "postDate": "2019-01-29T13:19:35.520Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 459815,
      "author_name": "Yifan Xie",
      "author_url": "",
      "post_date": "2019-01-22T12:20:42.643000",
      "content": "<p>I haven't check my best LB among the models in my stacking, but I estimate it brings about 0.005 improvement on LB.\nIn terms of CV, stacking bring about 0.012, but part of that is very likely to be overfitting. </p>\n\n<p>I think the improvement of stacking decreases as your score moves toward lower RMSE - I remember in my very first stacking test run with LGB model at LB3.695, I could achieve stacking score LB 3.685</p>",
      "votes": 4,
      "replies": []
    },
    {
      "id": 457372,
      "author_name": "AmirH",
      "author_url": "",
      "post_date": "2019-01-17T09:45:43.340000",
      "content": "<p>Very basic stacking got me 0.011 improvement.\nI will use my entire stacking and blending pipeline once i finish some more FE. \nMy single model is not as good as those i see here </p>",
      "votes": 1,
      "replies": [
        {
          "id": 457389,
          "author_name": "Manoj",
          "author_url": "",
          "post_date": "2019-01-17T10:19:00.193000",
          "content": "<p>thanks, just wanted to have an idea as i am not able to improve my single model\ngood to know that there is a decent scope in stacking :)</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 459787,
          "author_name": "Jianfei Wang",
          "author_url": "",
          "post_date": "2019-01-22T11:19:58.107000",
          "content": "<p>0.011 imp is very a very big imp for this competition.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 459820,
          "author_name": "AmirH",
          "author_url": "",
          "post_date": "2019-01-22T12:31:11.207000",
          "content": "<p>Thanks Wang,\nI am now working on my entire pipeline, in order to run an entire solution incl. blending and stacking, and then with new LB standing - look for a teammate with strong single model.\nAlso - my CV-LB relationship remains relatively consistent with stacking.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 463951,
          "author_name": "dylonLL",
          "author_url": "",
          "post_date": "2019-01-31T00:13:17.267000",
          "content": "<p>my problem is that my XGB model is doing so bad that it worsen my blending</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 457230,
      "author_name": "Marcus Lin",
      "author_url": "",
      "post_date": "2019-01-17T05:38:24.180000",
      "content": "<p>No idea, nothing successful for me.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 457280,
          "author_name": "Manoj",
          "author_url": "",
          "post_date": "2019-01-17T07:22:10.533000",
          "content": "<p>current LB is your single model , wow !</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 457353,
          "author_name": "Marcus Lin",
          "author_url": "",
          "post_date": "2019-01-17T09:10:01.750000",
          "content": "<p>No, we use post-process trick, but we definitely won't use it for final submissions.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 463124,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-01-29T13:19:35.520000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "457213": "How much extra are you able to achieve on the LB using stacking and blending over your best single model ?\n\nEdit 1: \nin my case  LB 3.689 (best single model) ---&gt;&gt;  3.686 \n\nEdit 2:\n LB 3.685 (best single model) ---&gt;&gt;  3.674 (done by teammate @AmirH)",
    "459815": "I haven't check my best LB among the models in my stacking, but I estimate it brings about 0.005 improvement on LB.\nIn terms of CV, stacking bring about 0.012, but part of that is very likely to be overfitting. \n\nI think the improvement of stacking decreases as your score moves toward lower RMSE - I remember in my very first stacking test run with LGB model at LB3.695, I could achieve stacking score LB 3.685\n",
    "457372": "Very basic stacking got me 0.011 improvement.\nI will use my entire stacking and blending pipeline once i finish some more FE. \nMy single model is not as good as those i see here ",
    "457230": "No idea, nothing successful for me.",
    "463124": ""
  }
}