{
  "id": 106276,
  "title": "Approach around LB 470th.",
  "url": "/competitions/champs-scalar-coupling/discussion/106276",
  "author_name": "",
  "post_date": "2019-08-29T00:58:28.025150100Z",
  "votes": 5,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Competition ended.  Congrats all people who get medal and knowledge.\nI didn't get good score, and there're(will be) many nice score's approaches.  So, this discussion is made for beginners like me.</p>\n\n<p>Validation:\nI just simply made models for each types and optimize MAE.  Just mean of these scores lead good correlation between CV and LB.  Scores's differences are just in +-0.05.\nThere's a trick people said (from discussion info), but I didn't notice.</p>\n\n<p>EDA:\nI didn't do except kernels and discussions.  I think it was bad.</p>\n\n<p>Feature Engineering:\nAll from kernels.  But, as the limits of memory, I made for each kernels and tuned models.  I think it bad because there's a people who get -1.9x(~10% score) said all features were made from discussions.\nMaking kernels for each types, and creating features and models were maybe good.</p>\n\n<p>Researching:\nI've read some informations about important features (distance, angle) and main approaches (MPNN, GNN), but I didn't understand and use it well.  Lack of experience of reading papers and scripts.</p>\n\n<p>Thanks for reading.</p>",
  "messages": [
    {
      "id": "610759",
      "postDate": "08/29/2019 00:58:28",
      "content": "<p>Competition ended.  Congrats all people who get medal and knowledge.\nI didn't get good score, and there're(will be) many nice score's approaches.  So, this discussion is made for beginners like me.</p>\n\n<p>Validation:\nI just simply made models for each types and optimize MAE.  Just mean of these scores lead good correlation between CV and LB.  Scores's differences are just in +-0.05.\nThere's a trick people said (from discussion info), but I didn't notice.</p>\n\n<p>EDA:\nI didn't do except kernels and discussions.  I think it was bad.</p>\n\n<p>Feature Engineering:\nAll from kernels.  But, as the limits of memory, I made for each kernels and tuned models.  I think it bad because there's a people who get -1.9x(~10% score) said all features were made from discussions.\nMaking kernels for each types, and creating features and models were maybe good.</p>\n\n<p>Researching:\nI've read some informations about important features (distance, angle) and main approaches (MPNN, GNN), but I didn't understand and use it well.  Lack of experience of reading papers and scripts.</p>\n\n<p>Thanks for reading.</p>",
      "rawMarkdown": "Competition ended.  Congrats all people who get medal and knowledge.\nI didn't get good score, and there're(will be) many nice score's approaches.  So, this discussion is made for beginners like me.\n\n\nValidation:\nI just simply made models for each types and optimize MAE.  Just mean of these scores lead good correlation between CV and LB.  Scores's differences are just in +-0.05.\nThere's a trick people said (from discussion info), but I didn't notice.\n\nEDA:\nI didn't do except kernels and discussions.  I think it was bad.\n\nFeature Engineering:\nAll from kernels.  But, as the limits of memory, I made for each kernels and tuned models.  I think it bad because there's a people who get -1.9x(~10% score) said all features were made from discussions.\nMaking kernels for each types, and creating features and models were maybe good.\n\nResearching:\nI've read some informations about important features (distance, angle) and main approaches (MPNN, GNN), but I didn't understand and use it well.  Lack of experience of reading papers and scripts.\n\nThanks for reading.",
      "votes": null
    },
    {
      "id": "610765",
      "postDate": "08/29/2019 01:01:33",
      "content": "<p>GNN was a problem, and last week without GPUs ... but my private LGBM is -2.2, all about features, we spent 1 morning finetuning using bayessian optimization, that's all :)</p>",
      "rawMarkdown": "GNN was a problem, and last week without GPUs ... but my private LGBM is -2.2, all about features, we spent 1 morning finetuning using bayessian optimization, that's all :)",
      "votes": null
    },
    {
      "id": "612711",
      "postDate": "08/30/2019 00:34:06",
      "content": "<p>Thanks for sharing, and congrats your medal.  LGBM with -2.2 is amazing for me.  Well, it was good to concentrate LGBM to concern about the learning/calculating cost of GNN ;)</p>",
      "rawMarkdown": "Thanks for sharing, and congrats your medal.  LGBM with -2.2 is amazing for me.  Well, it was good to concentrate LGBM to concern about the learning/calculating cost of GNN ;)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 610765,
      "author_name": "jesucristo",
      "author_url": "",
      "post_date": "08/29/2019 01:01:33",
      "content": "<p>GNN was a problem, and last week without GPUs ... but my private LGBM is -2.2, all about features, we spent 1 morning finetuning using bayessian optimization, that's all :)</p>",
      "votes": null,
      "replies": [
        {
          "id": 612711,
          "author_name": "jpn888",
          "author_url": "",
          "post_date": "08/30/2019 00:34:06",
          "content": "<p>Thanks for sharing, and congrats your medal.  LGBM with -2.2 is amazing for me.  Well, it was good to concentrate LGBM to concern about the learning/calculating cost of GNN ;)</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "610759": "Competition ended.  Congrats all people who get medal and knowledge.\nI didn't get good score, and there're(will be) many nice score's approaches.  So, this discussion is made for beginners like me.\n\n\nValidation:\nI just simply made models for each types and optimize MAE.  Just mean of these scores lead good correlation between CV and LB.  Scores's differences are just in +-0.05.\nThere's a trick people said (from discussion info), but I didn't notice.\n\nEDA:\nI didn't do except kernels and discussions.  I think it was bad.\n\nFeature Engineering:\nAll from kernels.  But, as the limits of memory, I made for each kernels and tuned models.  I think it bad because there's a people who get -1.9x(~10% score) said all features were made from discussions.\nMaking kernels for each types, and creating features and models were maybe good.\n\nResearching:\nI've read some informations about important features (distance, angle) and main approaches (MPNN, GNN), but I didn't understand and use it well.  Lack of experience of reading papers and scripts.\n\nThanks for reading.",
    "610765": "GNN was a problem, and last week without GPUs ... but my private LGBM is -2.2, all about features, we spent 1 morning finetuning using bayessian optimization, that's all :)",
    "612711": "Thanks for sharing, and congrats your medal.  LGBM with -2.2 is amazing for me.  Well, it was good to concentrate LGBM to concern about the learning/calculating cost of GNN ;)"
  },
  "source": "meta"
}