{
  "id": 420340,
  "title": "About the uselessness of features extraction in this competition",
  "url": "/competitions/tlvmc-parkinsons-freezing-gait-prediction/discussion/420340",
  "author_name": "Alberto Annoni",
  "post_date": "2023-06-30T10:24:37.376000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I'm really impressed by the top solutions.<br>\nWhen I first saw the dataset I was sure that applying Fourier Transform passing to the frequency domain would be crucial for good results.<br>\nWhen I saw the top solutions I was shocked that this approach was totally absent and most solution used plain raw acceleration data, someone also managing to fit a single model for both tdcs and de dataset (which was my secret dream).<br>\nSomeone have an explanation for that? <br>\nModels are nowadays so good to need minimal features extraction in the time-series context?<br>\nDo you agree that top performer models were either variations of residual networks or encoder + attention architectures?</p>",
  "messages": [
    {
      "id": 2324017,
      "postDate": "2023-06-30T10:24:37.377Z",
      "content": "<p>I'm really impressed by the top solutions.<br>\nWhen I first saw the dataset I was sure that applying Fourier Transform passing to the frequency domain would be crucial for good results.<br>\nWhen I saw the top solutions I was shocked that this approach was totally absent and most solution used plain raw acceleration data, someone also managing to fit a single model for both tdcs and de dataset (which was my secret dream).<br>\nSomeone have an explanation for that? <br>\nModels are nowadays so good to need minimal features extraction in the time-series context?<br>\nDo you agree that top performer models were either variations of residual networks or encoder + attention architectures?</p>",
      "rawMarkdown": "I'm really impressed by the top solutions.\nWhen I first saw the dataset I was sure that applying Fourier Transform passing to the frequency domain would be crucial for good results.\nWhen I saw the top solutions I was shocked that this approach was totally absent and most solution used plain raw acceleration data, someone also managing to fit a single model for both tdcs and de dataset (which was my secret dream).\nSomeone have an explanation for that? \nModels are nowadays so good to need minimal features extraction in the time-series context?\nDo you agree that top performer models were either variations of residual networks or encoder + attention architectures?",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2324017": "I'm really impressed by the top solutions.\nWhen I first saw the dataset I was sure that applying Fourier Transform passing to the frequency domain would be crucial for good results.\nWhen I saw the top solutions I was shocked that this approach was totally absent and most solution used plain raw acceleration data, someone also managing to fit a single model for both tdcs and de dataset (which was my secret dream).\nSomeone have an explanation for that? \nModels are nowadays so good to need minimal features extraction in the time-series context?\nDo you agree that top performer models were either variations of residual networks or encoder + attention architectures?"
  }
}