{
  "id": 507073,
  "title": "Can we extract and use lat/lon/time information",
  "url": "/competitions/leap-atmospheric-physics-ai-climsim/discussion/507073",
  "author_name": "Renu Singh",
  "post_date": "2024-05-24T10:08:28.843000",
  "votes": 0,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Related to the topic about using lat/lon/time info: <a href=\"https://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/495258\" target=\"_blank\">https://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/495258</a></p>\n<p>Can we develop a model that is aware of spatial and/or temporal info?</p>\n<p>And if so, can we recover the lat/lon/time information from the test set by assuming that the 384 locations are in order and every 7th timestep was subsampled just like the train set?</p>",
  "messages": [
    {
      "id": 2833728,
      "postDate": "2024-05-24T11:23:27.793Z",
      "content": "<p>'assuming that the 384 locations are in order'- Maybe. May be worth a shot to verify by comparing means  of features of supposed test grid to means of features of grids locations in training set.<br>\n'every 7th timestep was subsampled just like the train set' - No. If you look at running mean of features, you see that the duration of test set is longer than supposed duration. Several (a lot) of features exhibit cyclic pattern with a cycle per year, and test set's ~600K should be less than a year, but you will see a pattern that suggest it is longer. However, it is very easy to recover the 'stretching' and 'displacement' constants by comparing the cyclic patterns.</p>",
      "rawMarkdown": "'assuming that the 384 locations are in order'- Maybe. May be worth a shot to verify by comparing means  of features of supposed test grid to means of features of grids locations in training set.\n'every 7th timestep was subsampled just like the train set' - No. If you look at running mean of features, you see that the duration of test set is longer than supposed duration. Several (a lot) of features exhibit cyclic pattern with a cycle per year, and test set's ~600K should be less than a year, but you will see a pattern that suggest it is longer. However, it is very easy to recover the 'stretching' and 'displacement' constants by comparing the cyclic patterns.",
      "votes": 1,
      "replies": [
        {
          "id": 2833736,
          "postDate": "2024-05-24T11:29:12.937Z",
          "content": "<p>Regarding grid, I just made a quick check. We know that there are 384 grid locations. For train set 10091520/384 = 26280 which is good. But for test 625000/384 = 1627.60416667 so there is some trick. Several grid points were dropped for sure.</p>",
          "rawMarkdown": "Regarding grid, I just made a quick check. We know that there are 384 grid locations. For train set 10091520/384 = 26280 which is good. But for test 625000/384 = 1627.60416667 so there is some trick. Several grid points were dropped for sure."
        }
      ]
    },
    {
      "id": 2833609,
      "postDate": "2024-05-24T10:08:28.843Z",
      "content": "<p>Related to the topic about using lat/lon/time info: <a href=\"https://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/495258\" target=\"_blank\">https://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/495258</a></p>\n<p>Can we develop a model that is aware of spatial and/or temporal info?</p>\n<p>And if so, can we recover the lat/lon/time information from the test set by assuming that the 384 locations are in order and every 7th timestep was subsampled just like the train set?</p>",
      "rawMarkdown": "Related to the topic about using lat/lon/time info: https://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/495258\n\nCan we develop a model that is aware of spatial and/or temporal info?\n\nAnd if so, can we recover the lat/lon/time information from the test set by assuming that the 384 locations are in order and every 7th timestep was subsampled just like the train set?"
    }
  ],
  "comments": [
    {
      "id": 2833728,
      "author_name": "greySnow",
      "author_url": "",
      "post_date": "2024-05-24T11:23:27.793000",
      "content": "<p>'assuming that the 384 locations are in order'- Maybe. May be worth a shot to verify by comparing means  of features of supposed test grid to means of features of grids locations in training set.<br>\n'every 7th timestep was subsampled just like the train set' - No. If you look at running mean of features, you see that the duration of test set is longer than supposed duration. Several (a lot) of features exhibit cyclic pattern with a cycle per year, and test set's ~600K should be less than a year, but you will see a pattern that suggest it is longer. However, it is very easy to recover the 'stretching' and 'displacement' constants by comparing the cyclic patterns.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 2833736,
          "author_name": "greySnow",
          "author_url": "",
          "post_date": "2024-05-24T11:29:12.937000",
          "content": "<p>Regarding grid, I just made a quick check. We know that there are 384 grid locations. For train set 10091520/384 = 26280 which is good. But for test 625000/384 = 1627.60416667 so there is some trick. Several grid points were dropped for sure.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2833728": "'assuming that the 384 locations are in order'- Maybe. May be worth a shot to verify by comparing means  of features of supposed test grid to means of features of grids locations in training set.\n'every 7th timestep was subsampled just like the train set' - No. If you look at running mean of features, you see that the duration of test set is longer than supposed duration. Several (a lot) of features exhibit cyclic pattern with a cycle per year, and test set's ~600K should be less than a year, but you will see a pattern that suggest it is longer. However, it is very easy to recover the 'stretching' and 'displacement' constants by comparing the cyclic patterns.",
    "2833609": "Related to the topic about using lat/lon/time info: https://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/495258\n\nCan we develop a model that is aware of spatial and/or temporal info?\n\nAnd if so, can we recover the lat/lon/time information from the test set by assuming that the 384 locations are in order and every 7th timestep was subsampled just like the train set?"
  }
}