{
  "id": 449017,
  "title": "Preparation of train data for RNN models",
  "url": "/competitions/child-mind-institute-detect-sleep-states/discussion/449017",
  "author_name": "",
  "post_date": "2023-10-22T17:09:06.580952700Z",
  "votes": 6,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I was thinking about which way is better to handle this task:</p>\n<ol>\n<li>downsample (aggregate) the train series by some factor, and use aggregated mean, std, etc as new features;</li>\n<li>use rolling window - split the train series into N overlapping blocks with window length W and stride S, and use resulting series with features like lag, cumsum.</li>\n</ol>\n<p>First approach was used by <a href=\"https://www.kaggle.com/werus23\" target=\"_blank\">@werus23</a> is his notebooks, and I managed to achieve my currently best result (LB 0.627) with it. Later I tried second approach, and my currently best result  is LB 0.609. Also I noticed that larger window length/stride during an inference radically improves LB score (by ~0.05 - 0.1).</p>",
  "messages": [
    {
      "id": "2492611",
      "postDate": "10/22/2023 17:09:06",
      "content": "<p>I was thinking about which way is better to handle this task:</p>\n<ol>\n<li>downsample (aggregate) the train series by some factor, and use aggregated mean, std, etc as new features;</li>\n<li>use rolling window - split the train series into N overlapping blocks with window length W and stride S, and use resulting series with features like lag, cumsum.</li>\n</ol>\n<p>First approach was used by <a href=\"https://www.kaggle.com/werus23\" target=\"_blank\">@werus23</a> is his notebooks, and I managed to achieve my currently best result (LB 0.627) with it. Later I tried second approach, and my currently best result  is LB 0.609. Also I noticed that larger window length/stride during an inference radically improves LB score (by ~0.05 - 0.1).</p>",
      "rawMarkdown": "I was thinking about which way is better to handle this task:\n1. downsample (aggregate) the train series by some factor, and use aggregated mean, std, etc as new features;\n2. use rolling window - split the train series into N overlapping blocks with window length W and stride S, and use resulting series with features like lag, cumsum.\n\nFirst approach was used by @werus23 is his notebooks, and I managed to achieve my currently best result (LB 0.627) with it. Later I tried second approach, and my currently best result  is LB 0.609. Also I noticed that larger window length/stride during an inference radically improves LB score (by ~0.05 - 0.1).",
      "votes": null
    },
    {
      "id": "2492795",
      "postDate": "10/22/2023 21:57:44",
      "content": "<p>Thanks for sharing <a href=\"https://www.kaggle.com/atamazian\" target=\"_blank\">@atamazian</a>. I have only tried ML so far, but will LB become NN &gt; ML?</p>",
      "rawMarkdown": "Thanks for sharing @atamazian. I have only tried ML so far, but will LB become NN > ML?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2492795,
      "author_name": "mitsuyasuhoshino",
      "author_url": "",
      "post_date": "10/22/2023 21:57:44",
      "content": "<p>Thanks for sharing <a href=\"https://www.kaggle.com/atamazian\" target=\"_blank\">@atamazian</a>. I have only tried ML so far, but will LB become NN &gt; ML?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2492611": "I was thinking about which way is better to handle this task:\n1. downsample (aggregate) the train series by some factor, and use aggregated mean, std, etc as new features;\n2. use rolling window - split the train series into N overlapping blocks with window length W and stride S, and use resulting series with features like lag, cumsum.\n\nFirst approach was used by @werus23 is his notebooks, and I managed to achieve my currently best result (LB 0.627) with it. Later I tried second approach, and my currently best result  is LB 0.609. Also I noticed that larger window length/stride during an inference radically improves LB score (by ~0.05 - 0.1).",
    "2492795": "Thanks for sharing @atamazian. I have only tried ML so far, but will LB become NN > ML?"
  },
  "source": "meta"
}