{
  "id": 94374,
  "title": "Virtual LB 130th: Mel spectrogram feature + class balance simple solution",
  "url": "/competitions/LANL-Earthquake-Prediction/discussion/94374",
  "author_name": "daisukelab",
  "post_date": "2019-06-04T06:36:41.930000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p><a href=\"https://www.kaggle.com/daisukelab/lanl-solution-by-mel-spectrogram-dataset-2\">https://www.kaggle.com/daisukelab/lanl-solution-by-mel-spectrogram-dataset-2</a>\nSharing an approach which people might be interested.</p>\n\n<h3>Feature generation</h3>\n\n<ul>\n<li>Once converted into 62,900 mel-spectrogram data by fetching source 150,000 samples with sliding window.\n<ul><li>fs=6MHz, step=40000/34, n_fft=40000 =&gt; mel-spectrogram data shape=[62900, 128, 128]</li></ul></li>\n<li>Convert mel-spectrogram data into 3 aggregated features: mean, std, trend like. =&gt; data shape=[62900, 384]</li>\n</ul>\n\n<h3>Balancing dataset</h3>\n\n<ul>\n<li>Apply stratified class balancing described in my kernel <a href=\"https://www.kaggle.com/daisukelab/balancing-dataset-while-earthquake-is-happening\">https://www.kaggle.com/daisukelab/balancing-dataset-while-earthquake-is-happening</a></li>\n<li>By virtually handing dataset as labeled set.</li>\n</ul>\n\n<h3>Model</h3>\n\n<ul>\n<li>Simple LGBM.</li>\n</ul>\n\n<p>Unfortunately I have deleted mel spectrogram dataset, I can post it upon your request later.</p>",
  "messages": [
    {
      "id": 542856,
      "postDate": "2019-06-04T06:36:41.930Z",
      "content": "<p><a href=\"https://www.kaggle.com/daisukelab/lanl-solution-by-mel-spectrogram-dataset-2\">https://www.kaggle.com/daisukelab/lanl-solution-by-mel-spectrogram-dataset-2</a>\nSharing an approach which people might be interested.</p>\n\n<h3>Feature generation</h3>\n\n<ul>\n<li>Once converted into 62,900 mel-spectrogram data by fetching source 150,000 samples with sliding window.\n<ul><li>fs=6MHz, step=40000/34, n_fft=40000 =&gt; mel-spectrogram data shape=[62900, 128, 128]</li></ul></li>\n<li>Convert mel-spectrogram data into 3 aggregated features: mean, std, trend like. =&gt; data shape=[62900, 384]</li>\n</ul>\n\n<h3>Balancing dataset</h3>\n\n<ul>\n<li>Apply stratified class balancing described in my kernel <a href=\"https://www.kaggle.com/daisukelab/balancing-dataset-while-earthquake-is-happening\">https://www.kaggle.com/daisukelab/balancing-dataset-while-earthquake-is-happening</a></li>\n<li>By virtually handing dataset as labeled set.</li>\n</ul>\n\n<h3>Model</h3>\n\n<ul>\n<li>Simple LGBM.</li>\n</ul>\n\n<p>Unfortunately I have deleted mel spectrogram dataset, I can post it upon your request later.</p>",
      "rawMarkdown": "https://www.kaggle.com/daisukelab/lanl-solution-by-mel-spectrogram-dataset-2\nSharing an approach which people might be interested.\n\n### Feature generation\n- Once converted into 62,900 mel-spectrogram data by fetching source 150,000 samples with sliding window.\n    - fs=6MHz, step=40000/34, n_fft=40000 =&gt; mel-spectrogram data shape=[62900, 128, 128]\n- Convert mel-spectrogram data into 3 aggregated features: mean, std, trend like. =&gt; data shape=[62900, 384]\n\n### Balancing dataset\n- Apply stratified class balancing described in my kernel https://www.kaggle.com/daisukelab/balancing-dataset-while-earthquake-is-happening\n- By virtually handing dataset as labeled set.\n\n### Model\n- Simple LGBM.\n\nUnfortunately I have deleted mel spectrogram dataset, I can post it upon your request later.",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "542856": "https://www.kaggle.com/daisukelab/lanl-solution-by-mel-spectrogram-dataset-2\nSharing an approach which people might be interested.\n\n### Feature generation\n- Once converted into 62,900 mel-spectrogram data by fetching source 150,000 samples with sliding window.\n    - fs=6MHz, step=40000/34, n_fft=40000 =&gt; mel-spectrogram data shape=[62900, 128, 128]\n- Convert mel-spectrogram data into 3 aggregated features: mean, std, trend like. =&gt; data shape=[62900, 384]\n\n### Balancing dataset\n- Apply stratified class balancing described in my kernel https://www.kaggle.com/daisukelab/balancing-dataset-while-earthquake-is-happening\n- By virtually handing dataset as labeled set.\n\n### Model\n- Simple LGBM.\n\nUnfortunately I have deleted mel spectrogram dataset, I can post it upon your request later."
  }
}