{
  "id": 372800,
  "title": "Which is best?ML(xgb,lgb) or DL(deepsort,yolo...)?",
  "url": "/competitions/nfl-player-contact-detection/discussion/372800",
  "author_name": "",
  "post_date": "2022-12-18T06:02:50.391422700Z",
  "votes": 6,
  "comment_count": 7,
  "views": 0,
  "content": "<p>My cv is 0.669,lb 0.658(xgb model).I tried many FE and tricks,but I think it is hard for me to improve my lb score above 0.7+.</p>",
  "messages": [
    {
      "id": "2068599",
      "postDate": "12/18/2022 06:02:50",
      "content": "<p>My cv is 0.669,lb 0.658(xgb model).I tried many FE and tricks,but I think it is hard for me to improve my lb score above 0.7+.</p>",
      "rawMarkdown": "My cv is 0.669,lb 0.658(xgb model).I tried many FE and tricks,but I think it is hard for me to improve my lb score above 0.7+.",
      "votes": null
    },
    {
      "id": "2068646",
      "postDate": "12/18/2022 07:39:09",
      "content": "<p>No such thing as \"best\". It depends on the feature space, task, hardware, etc, and in real life even maintainability comes into play.</p>",
      "rawMarkdown": "No such thing as \"best\". It depends on the feature space, task, hardware, etc, and in real life even maintainability comes into play.",
      "votes": null
    },
    {
      "id": "2069708",
      "postDate": "12/19/2022 09:04:13",
      "content": "<p>I think video data(mp4) has more useful information than csv data(helmets.csv,tracking.csv),so deepsort and yolov5 can get higher score in this competition.What' more,good features are hard to find for tree model(xgb,lgb).</p>",
      "rawMarkdown": "I think video data(mp4) has more useful information than csv data(helmets.csv,tracking.csv),so deepsort and yolov5 can get higher score in this competition.What' more,good features are hard to find for tree model(xgb,lgb).",
      "votes": null
    },
    {
      "id": "2070021",
      "postDate": "12/19/2022 14:39:15",
      "content": "<p>I see your point, sure, for the image data, I dont think GBDT stands a chance.<br>\nAlso, definition of \"best\" changes depending on context, if all we care is raw performance, its a much easier question to answer</p>",
      "rawMarkdown": "I see your point, sure, for the image data, I dont think GBDT stands a chance.\nAlso, definition of \"best\" changes depending on context, if all we care is raw performance, its a much easier question to answer",
      "votes": null
    },
    {
      "id": "2070057",
      "postDate": "12/19/2022 15:06:27",
      "content": "<p>Yeah.You are top 3 now，good luck to you！</p>",
      "rawMarkdown": "Yeah.You are top 3 now，good luck to you！",
      "votes": null
    },
    {
      "id": "2076960",
      "postDate": "12/27/2022 03:51:47",
      "content": "<p><a href=\"https://www.kaggle.com/cjzccc\" target=\"_blank\">@cjzccc</a> what types of features were you able to create to get to a 0.669 cv score? Were most of them from the tracking data, or did you find the helmet data to be most impactful?</p>",
      "rawMarkdown": "cjzccc what types of features were you able to create to get to a 0.669 cv score? Were most of them from the tracking data, or did you find the helmet data to be most impactful?",
      "votes": null
    },
    {
      "id": "2078005",
      "postDate": "12/28/2022 02:51:15",
      "content": "<p>Most features are from tracking data.</p>",
      "rawMarkdown": "Most features are from tracking data.",
      "votes": null
    },
    {
      "id": "2078577",
      "postDate": "12/28/2022 12:44:00",
      "content": "<p>Awesome! It's great to know it's possible to get a really strong score with a tree model.</p>",
      "rawMarkdown": "Awesome! It's great to know it's possible to get a really strong score with a tree model.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2068646,
      "author_name": "carloshuertas",
      "author_url": "",
      "post_date": "12/18/2022 07:39:09",
      "content": "<p>No such thing as \"best\". It depends on the feature space, task, hardware, etc, and in real life even maintainability comes into play.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2069708,
      "author_name": "cjzccc",
      "author_url": "",
      "post_date": "12/19/2022 09:04:13",
      "content": "<p>I think video data(mp4) has more useful information than csv data(helmets.csv,tracking.csv),so deepsort and yolov5 can get higher score in this competition.What' more,good features are hard to find for tree model(xgb,lgb).</p>",
      "votes": null,
      "replies": [
        {
          "id": 2070021,
          "author_name": "carloshuertas",
          "author_url": "",
          "post_date": "12/19/2022 14:39:15",
          "content": "<p>I see your point, sure, for the image data, I dont think GBDT stands a chance.<br>\nAlso, definition of \"best\" changes depending on context, if all we care is raw performance, its a much easier question to answer</p>",
          "votes": null,
          "replies": [
            {
              "id": 2070057,
              "author_name": "cjzccc",
              "author_url": "",
              "post_date": "12/19/2022 15:06:27",
              "content": "<p>Yeah.You are top 3 now，good luck to you！</p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2076960,
      "author_name": "ryancaldwell",
      "author_url": "",
      "post_date": "12/27/2022 03:51:47",
      "content": "<p><a href=\"https://www.kaggle.com/cjzccc\" target=\"_blank\">@cjzccc</a> what types of features were you able to create to get to a 0.669 cv score? Were most of them from the tracking data, or did you find the helmet data to be most impactful?</p>",
      "votes": null,
      "replies": [
        {
          "id": 2078005,
          "author_name": "cjzccc",
          "author_url": "",
          "post_date": "12/28/2022 02:51:15",
          "content": "<p>Most features are from tracking data.</p>",
          "votes": null,
          "replies": [
            {
              "id": 2078577,
              "author_name": "ryancaldwell",
              "author_url": "",
              "post_date": "12/28/2022 12:44:00",
              "content": "<p>Awesome! It's great to know it's possible to get a really strong score with a tree model.</p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2068599": "My cv is 0.669,lb 0.658(xgb model).I tried many FE and tricks,but I think it is hard for me to improve my lb score above 0.7+.",
    "2068646": "No such thing as \"best\". It depends on the feature space, task, hardware, etc, and in real life even maintainability comes into play.",
    "2069708": "I think video data(mp4) has more useful information than csv data(helmets.csv,tracking.csv),so deepsort and yolov5 can get higher score in this competition.What' more,good features are hard to find for tree model(xgb,lgb).",
    "2070021": "I see your point, sure, for the image data, I dont think GBDT stands a chance.\nAlso, definition of \"best\" changes depending on context, if all we care is raw performance, its a much easier question to answer",
    "2070057": "Yeah.You are top 3 now，good luck to you！",
    "2076960": "cjzccc what types of features were you able to create to get to a 0.669 cv score? Were most of them from the tracking data, or did you find the helmet data to be most impactful?",
    "2078005": "Most features are from tracking data.",
    "2078577": "Awesome! It's great to know it's possible to get a really strong score with a tree model."
  },
  "source": "meta"
}