{
  "id": 215463,
  "title": "3rd rank solution",
  "url": "/competitions/predict-volcanic-eruptions-ingv-oe/discussion/215463",
  "author_name": "Hiroyuki",
  "post_date": "2021-01-30T00:03:05.006000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>First of all,  I’d like to thank INGV and Kaggle for hosting this competition.<br>\nThis is my first Kaggle competition, and I’m very happy to be third place.<br>\nAnd I apologize for being so late to post this to this discussion forum, and for not having my idea properly organized.</p>\n<p>During almost all the competition period, I struggled with the difference between valid scores and public scores.<br>\nAt last (2 days before the deadline), I noticed the test data had 2 types, one is similar to the train data (Type A) and the other is not (Type B).<br>\nUsing the adversarial validation, it clearly appeared. <br>\nI evaluated Type A and Type B separately.</p>\n<p><em>For Type A</em></p>\n<p>Features: </p>\n<ul>\n<li>frequency spectrums summarized by log scales</li>\n</ul>\n<p>Algorithm : KNN Regressor</p>\n<p><em>For Type B</em></p>\n<p>Features:</p>\n<ul>\n<li>frequency spectrums summarized by log scales</li>\n<li>variations(std, skew and kurtosis) of 0.8-3.5 Hz, and 3.5-8.0Hz frequency</li>\n<li>max, mean, std of  sensors’ values</li>\n</ul>\n<p>Algorithm : Light GBM</p>\n<p>I would be very happy to receive any advice or questions.</p>",
  "messages": [
    {
      "id": 1176984,
      "postDate": "2021-01-30T00:03:05.007Z",
      "content": "<p>First of all,  I’d like to thank INGV and Kaggle for hosting this competition.<br>\nThis is my first Kaggle competition, and I’m very happy to be third place.<br>\nAnd I apologize for being so late to post this to this discussion forum, and for not having my idea properly organized.</p>\n<p>During almost all the competition period, I struggled with the difference between valid scores and public scores.<br>\nAt last (2 days before the deadline), I noticed the test data had 2 types, one is similar to the train data (Type A) and the other is not (Type B).<br>\nUsing the adversarial validation, it clearly appeared. <br>\nI evaluated Type A and Type B separately.</p>\n<p><em>For Type A</em></p>\n<p>Features: </p>\n<ul>\n<li>frequency spectrums summarized by log scales</li>\n</ul>\n<p>Algorithm : KNN Regressor</p>\n<p><em>For Type B</em></p>\n<p>Features:</p>\n<ul>\n<li>frequency spectrums summarized by log scales</li>\n<li>variations(std, skew and kurtosis) of 0.8-3.5 Hz, and 3.5-8.0Hz frequency</li>\n<li>max, mean, std of  sensors’ values</li>\n</ul>\n<p>Algorithm : Light GBM</p>\n<p>I would be very happy to receive any advice or questions.</p>",
      "rawMarkdown": "First of all,  I’d like to thank INGV and Kaggle for hosting this competition.\nThis is my first Kaggle competition, and I’m very happy to be third place.\nAnd I apologize for being so late to post this to this discussion forum, and for not having my idea properly organized.\n\nDuring almost all the competition period, I struggled with the difference between valid scores and public scores.\nAt last (2 days before the deadline), I noticed the test data had 2 types, one is similar to the train data (Type A) and the other is not (Type B).\nUsing the adversarial validation, it clearly appeared. \nI evaluated Type A and Type B separately.\n\n*For Type A*\n\nFeatures: \n- frequency spectrums summarized by log scales\n\nAlgorithm : KNN Regressor\n\n\n*For Type B*\n\nFeatures:\n- frequency spectrums summarized by log scales\n- variations(std, skew and kurtosis) of 0.8-3.5 Hz, and 3.5-8.0Hz frequency\n- max, mean, std of  sensors’ values\n\nAlgorithm : Light GBM\n\nI would be very happy to receive any advice or questions.\n",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1176984": "First of all,  I’d like to thank INGV and Kaggle for hosting this competition.\nThis is my first Kaggle competition, and I’m very happy to be third place.\nAnd I apologize for being so late to post this to this discussion forum, and for not having my idea properly organized.\n\nDuring almost all the competition period, I struggled with the difference between valid scores and public scores.\nAt last (2 days before the deadline), I noticed the test data had 2 types, one is similar to the train data (Type A) and the other is not (Type B).\nUsing the adversarial validation, it clearly appeared. \nI evaluated Type A and Type B separately.\n\n*For Type A*\n\nFeatures: \n- frequency spectrums summarized by log scales\n\nAlgorithm : KNN Regressor\n\n\n*For Type B*\n\nFeatures:\n- frequency spectrums summarized by log scales\n- variations(std, skew and kurtosis) of 0.8-3.5 Hz, and 3.5-8.0Hz frequency\n- max, mean, std of  sensors’ values\n\nAlgorithm : Light GBM\n\nI would be very happy to receive any advice or questions.\n"
  }
}