{
  "id": 106328,
  "title": "Private LB -2.0 lgb features: Atom-Centered Symmetry Functions",
  "url": "/competitions/champs-scalar-coupling/discussion/106328",
  "author_name": "Shuhao Cao",
  "post_date": "2019-08-29T06:06:33.282000",
  "votes": 9,
  "comment_count": 1,
  "views": 0,
  "content": "<p><a href=\"https://www.kaggle.com/scaomath/lb-2-0-features-dscribe-acsf-descriptors\">https://www.kaggle.com/scaomath/lb-2-0-features-dscribe-acsf-descriptors</a></p>\n\n<p>The reference is of course:</p>\n\n<ul>\n<li><a href=\"https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors/notebook\">https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors/notebook</a></li>\n<li>(a seriously underrated kernel) <a href=\"https://www.kaggle.com/educatedguessing/diy-atom-centered-symmetry-functions\">https://www.kaggle.com/educatedguessing/diy-atom-centered-symmetry-functions</a></li>\n</ul>\n\n<p>I basically speed up the generation so I can play with the parameters within my <code>1JHN</code> feature playground.</p>\n\n<p>I generated a first set based on (1) parameters choices in the original paper, (2) permutation importance on <code>1JHN</code>, which has the fewest samples and has been my feature playground all 3 months.</p>\n\n<p>Then a second set using a new set of parameters were generated based on the feature selection of the first set, which eventually achieves -2.0 in private LB.</p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2745910%2Fd01a9e28e5ff93e86633c53b6323feb1%2FScreen%20Shot%202019-08-28%20at%2011.04.57%20PM.png?generation=1567058783847683&amp;alt=media\" alt=\"\"></p>",
  "messages": [
    {
      "id": 611144,
      "postDate": "2019-08-29T06:06:33.283Z",
      "content": "<p><a href=\"https://www.kaggle.com/scaomath/lb-2-0-features-dscribe-acsf-descriptors\">https://www.kaggle.com/scaomath/lb-2-0-features-dscribe-acsf-descriptors</a></p>\n\n<p>The reference is of course:</p>\n\n<ul>\n<li><a href=\"https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors/notebook\">https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors/notebook</a></li>\n<li>(a seriously underrated kernel) <a href=\"https://www.kaggle.com/educatedguessing/diy-atom-centered-symmetry-functions\">https://www.kaggle.com/educatedguessing/diy-atom-centered-symmetry-functions</a></li>\n</ul>\n\n<p>I basically speed up the generation so I can play with the parameters within my <code>1JHN</code> feature playground.</p>\n\n<p>I generated a first set based on (1) parameters choices in the original paper, (2) permutation importance on <code>1JHN</code>, which has the fewest samples and has been my feature playground all 3 months.</p>\n\n<p>Then a second set using a new set of parameters were generated based on the feature selection of the first set, which eventually achieves -2.0 in private LB.</p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2745910%2Fd01a9e28e5ff93e86633c53b6323feb1%2FScreen%20Shot%202019-08-28%20at%2011.04.57%20PM.png?generation=1567058783847683&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "https://www.kaggle.com/scaomath/lb-2-0-features-dscribe-acsf-descriptors\n\nThe reference is of course:\n\n- https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors/notebook\n- (a seriously underrated kernel) https://www.kaggle.com/educatedguessing/diy-atom-centered-symmetry-functions\n\nI basically speed up the generation so I can play with the parameters within my `1JHN` feature playground.\n\nI generated a first set based on (1) parameters choices in the original paper, (2) permutation importance on `1JHN`, which has the fewest samples and has been my feature playground all 3 months.\n\nThen a second set using a new set of parameters were generated based on the feature selection of the first set, which eventually achieves -2.0 in private LB.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2745910%2Fd01a9e28e5ff93e86633c53b6323feb1%2FScreen%20Shot%202019-08-28%20at%2011.04.57%20PM.png?generation=1567058783847683&amp;alt=media)\n",
      "votes": 8
    },
    {
      "id": 611343,
      "postDate": "2019-08-29T08:09:29.933Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 611343,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-08-29T08:09:29.933000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "611144": "https://www.kaggle.com/scaomath/lb-2-0-features-dscribe-acsf-descriptors\n\nThe reference is of course:\n\n- https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors/notebook\n- (a seriously underrated kernel) https://www.kaggle.com/educatedguessing/diy-atom-centered-symmetry-functions\n\nI basically speed up the generation so I can play with the parameters within my `1JHN` feature playground.\n\nI generated a first set based on (1) parameters choices in the original paper, (2) permutation importance on `1JHN`, which has the fewest samples and has been my feature playground all 3 months.\n\nThen a second set using a new set of parameters were generated based on the feature selection of the first set, which eventually achieves -2.0 in private LB.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2745910%2Fd01a9e28e5ff93e86633c53b6323feb1%2FScreen%20Shot%202019-08-28%20at%2011.04.57%20PM.png?generation=1567058783847683&amp;alt=media)\n",
    "611343": ""
  }
}