{
  "id": 545825,
  "title": "using autoencoder to encode time series data? :)",
  "url": "/competitions/child-mind-institute-problematic-internet-use/discussion/545825",
  "author_name": "",
  "post_date": "2024-11-12T10:52:22.688380300Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I've noticed that some public implementations use an autoencoder-based approach for encoding time series data. However, I have a few questions: 1) Why do they apply a standard scaler and a sigmoid output layer? 2) Why are they training on both the training and test sets? Is there a specific reason behind this approach, or am I missing something!</p>",
  "messages": [
    {
      "id": "3043342",
      "postDate": "11/12/2024 10:52:22",
      "content": "<p>I've noticed that some public implementations use an autoencoder-based approach for encoding time series data. However, I have a few questions: 1) Why do they apply a standard scaler and a sigmoid output layer? 2) Why are they training on both the training and test sets? Is there a specific reason behind this approach, or am I missing something!</p>",
      "rawMarkdown": "I've noticed that some public implementations use an autoencoder-based approach for encoding time series data. However, I have a few questions: 1) Why do they apply a standard scaler and a sigmoid output layer? 2) Why are they training on both the training and test sets? Is there a specific reason behind this approach, or am I missing something!",
      "votes": null
    },
    {
      "id": "3047369",
      "postDate": "11/16/2024 15:51:02",
      "content": "<p>I asked myself the same question. For 1) I don't know. Maybe some of the data has a normal distribution, haven't looked into it yet.<br>\n2) I think this is a calculated gamble, since the test data might have more time series data than the training data.<br>\nI changed it though.</p>",
      "rawMarkdown": "I asked myself the same question. For 1) I don't know. Maybe some of the data has a normal distribution, haven't looked into it yet.\n2) I think this is a calculated gamble, since the test data might have more time series data than the training data.\nI changed it though.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3047369,
      "author_name": "mariusheuser",
      "author_url": "",
      "post_date": "11/16/2024 15:51:02",
      "content": "<p>I asked myself the same question. For 1) I don't know. Maybe some of the data has a normal distribution, haven't looked into it yet.<br>\n2) I think this is a calculated gamble, since the test data might have more time series data than the training data.<br>\nI changed it though.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3043342": "I've noticed that some public implementations use an autoencoder-based approach for encoding time series data. However, I have a few questions: 1) Why do they apply a standard scaler and a sigmoid output layer? 2) Why are they training on both the training and test sets? Is there a specific reason behind this approach, or am I missing something!",
    "3047369": "I asked myself the same question. For 1) I don't know. Maybe some of the data has a normal distribution, haven't looked into it yet.\n2) I think this is a calculated gamble, since the test data might have more time series data than the training data.\nI changed it though."
  },
  "source": "meta"
}