{
  "id": 239998,
  "title": "14th some findings and solution",
  "url": "/competitions/indoor-location-navigation/writeups/ganchan-14th-some-findings-and-solution",
  "author_name": "",
  "post_date": "2021-05-18T09:22:33.618352600Z",
  "votes": 12,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I'd like to thank this great competition's hosts and congrats winners and all medalists!</p>\n<p>I think there are so many approach to this problem and I'm busy to read and understand other participants' awesome solusions.</p>\n<p>I also write a part of my solution.</p>\n<ol>\n<li><p>Findings and preprocessing<br>\nThere are much more wifi timestamp than waypoints and it is effective to use all wifi data (interpolating waypoint data).<br>\nSome BSSIDs show very high-correlated RSSI and getting rid of it is effective.<br>\nThere some MAC ADRESS for one ibeacon(uuid_major_minor) and we need to use MAC ADRESS instead of uuid.</p></li>\n<li><p>Models<br>\nI used LGBM to predict each point indipendently.<br>\nIn this model, train and predict the waypoints where wifi data exists and predict original WAYPOINT by adding relative posision data.<br>\nI used both wifi and ibeacon data.<br>\nI used only BSSID and MACADDR which exists both train and test.<br>\nThese data are transformed to rank.(not rssi value)</p></li>\n<li><p>Postprocessing<br>\nCost minimization with leakage and snap to grid.<br>\nIn cost minimization, I added leakage term into the original cost function.<br>\nIn snap to grid, I considered to relative positions (from sensor data) when select the proper grid point.<br>\nThe notebook is published <a href=\"https://www.kaggle.com/iwatatakuya/snap-to-grid-with-relative-position-data\" target=\"_blank\">here</a>.<br>\nIn my case, this postprocesing method improved score very much.</p></li>\n<li><p>Ensemble<br>\nEnsembles some models and postprocessing.</p></li>\n</ol>\n<p>Thank you for reading this far.<br>\nQuestions are welcome!</p>",
  "messages": [
    {
      "id": "1312830",
      "postDate": "05/18/2021 09:22:33",
      "content": "<p>I'd like to thank this great competition's hosts and congrats winners and all medalists!</p>\n<p>I think there are so many approach to this problem and I'm busy to read and understand other participants' awesome solusions.</p>\n<p>I also write a part of my solution.</p>\n<ol>\n<li><p>Findings and preprocessing<br>\nThere are much more wifi timestamp than waypoints and it is effective to use all wifi data (interpolating waypoint data).<br>\nSome BSSIDs show very high-correlated RSSI and getting rid of it is effective.<br>\nThere some MAC ADRESS for one ibeacon(uuid_major_minor) and we need to use MAC ADRESS instead of uuid.</p></li>\n<li><p>Models<br>\nI used LGBM to predict each point indipendently.<br>\nIn this model, train and predict the waypoints where wifi data exists and predict original WAYPOINT by adding relative posision data.<br>\nI used both wifi and ibeacon data.<br>\nI used only BSSID and MACADDR which exists both train and test.<br>\nThese data are transformed to rank.(not rssi value)</p></li>\n<li><p>Postprocessing<br>\nCost minimization with leakage and snap to grid.<br>\nIn cost minimization, I added leakage term into the original cost function.<br>\nIn snap to grid, I considered to relative positions (from sensor data) when select the proper grid point.<br>\nThe notebook is published <a href=\"https://www.kaggle.com/iwatatakuya/snap-to-grid-with-relative-position-data\" target=\"_blank\">here</a>.<br>\nIn my case, this postprocesing method improved score very much.</p></li>\n<li><p>Ensemble<br>\nEnsembles some models and postprocessing.</p></li>\n</ol>\n<p>Thank you for reading this far.<br>\nQuestions are welcome!</p>",
      "rawMarkdown": "I'd like to thank this great competition's hosts and congrats winners and all medalists!\n\nI think there are so many approach to this problem and I'm busy to read and understand other participants' awesome solusions.\n\nI also write a part of my solution.\n\n\n1. Findings and preprocessing\nThere are much more wifi timestamp than waypoints and it is effective to use all wifi data (interpolating waypoint data).\nSome BSSIDs show very high-correlated RSSI and getting rid of it is effective.\nThere some MAC ADRESS for one ibeacon(uuid_major_minor) and we need to use MAC ADRESS instead of uuid.\n\n2. Models\nI used LGBM to predict each point indipendently.\nIn this model, train and predict the waypoints where wifi data exists and predict original WAYPOINT by adding relative posision data.\nI used both wifi and ibeacon data.\nI used only BSSID and MACADDR which exists both train and test.\nThese data are transformed to rank.(not rssi value)\n\n3. Postprocessing\nCost minimization with leakage and snap to grid.\nIn cost minimization, I added leakage term into the original cost function.\nIn snap to grid, I considered to relative positions (from sensor data) when select the proper grid point.\nThe notebook is published [here](https://www.kaggle.com/iwatatakuya/snap-to-grid-with-relative-position-data).\nIn my case, this postprocesing method improved score very much.\n\n4. Ensemble\nEnsembles some models and postprocessing.\n\nThank you for reading this far.\nQuestions are welcome!",
      "votes": null
    },
    {
      "id": "1450831",
      "postDate": "08/05/2021 07:28:16",
      "content": "<p>Thanks for sharing</p>",
      "rawMarkdown": "Thanks for sharing",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1450831,
      "author_name": "staswinner",
      "author_url": "",
      "post_date": "08/05/2021 07:28:16",
      "content": "<p>Thanks for sharing</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1312830": "I'd like to thank this great competition's hosts and congrats winners and all medalists!\n\nI think there are so many approach to this problem and I'm busy to read and understand other participants' awesome solusions.\n\nI also write a part of my solution.\n\n\n1. Findings and preprocessing\nThere are much more wifi timestamp than waypoints and it is effective to use all wifi data (interpolating waypoint data).\nSome BSSIDs show very high-correlated RSSI and getting rid of it is effective.\nThere some MAC ADRESS for one ibeacon(uuid_major_minor) and we need to use MAC ADRESS instead of uuid.\n\n2. Models\nI used LGBM to predict each point indipendently.\nIn this model, train and predict the waypoints where wifi data exists and predict original WAYPOINT by adding relative posision data.\nI used both wifi and ibeacon data.\nI used only BSSID and MACADDR which exists both train and test.\nThese data are transformed to rank.(not rssi value)\n\n3. Postprocessing\nCost minimization with leakage and snap to grid.\nIn cost minimization, I added leakage term into the original cost function.\nIn snap to grid, I considered to relative positions (from sensor data) when select the proper grid point.\nThe notebook is published [here](https://www.kaggle.com/iwatatakuya/snap-to-grid-with-relative-position-data).\nIn my case, this postprocesing method improved score very much.\n\n4. Ensemble\nEnsembles some models and postprocessing.\n\nThank you for reading this far.\nQuestions are welcome!",
    "1450831": "Thanks for sharing"
  },
  "source": "meta"
}