{
  "id": 294677,
  "title": "PCA on ResNet Features",
  "url": "/competitions/cs5785-fall-2021-final/discussion/294677",
  "author_name": "",
  "post_date": "2021-12-12T05:05:24.381071Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Adding on to <a href=\"https://www.kaggle.com/c/cs5785-fall-2021-final/discussion/293537\" target=\"_blank\">Aarushi's post</a>, I did PCA for n_components = 50 - 200 and ran a ridge regression on each of them. I found out that lower number of features yields better results for validation. </p>\n<p><img src=\"https://i.postimg.cc/bNyvpLnP/Screen-Shot-2021-12-11-at-11-51-59-PM.png\" alt=\"\"></p>\n<p>Hope this is helpful! 😊</p>",
  "messages": [
    {
      "id": "1615275",
      "postDate": "12/12/2021 05:05:24",
      "content": "<p>Adding on to <a href=\"https://www.kaggle.com/c/cs5785-fall-2021-final/discussion/293537\" target=\"_blank\">Aarushi's post</a>, I did PCA for n_components = 50 - 200 and ran a ridge regression on each of them. I found out that lower number of features yields better results for validation. </p>\n<p><img src=\"https://i.postimg.cc/bNyvpLnP/Screen-Shot-2021-12-11-at-11-51-59-PM.png\" alt=\"\"></p>\n<p>Hope this is helpful! 😊</p>",
      "rawMarkdown": "Adding on to [Aarushi's post](https://www.kaggle.com/c/cs5785-fall-2021-final/discussion/293537), I did PCA for n_components = 50 - 200 and ran a ridge regression on each of them. I found out that lower number of features yields better results for validation. \n\n![](https://i.postimg.cc/bNyvpLnP/Screen-Shot-2021-12-11-at-11-51-59-PM.png)\n\nHope this is helpful! 😊",
      "votes": null
    },
    {
      "id": "1617294",
      "postDate": "12/14/2021 02:32:11",
      "content": "<p>Commenting on this in case anyone else is working. Our team is using a neural net and when fitting/validating just on the training data less features do better than more features. However, when we fitted to the whole training set and submitted to kaggle, we actually found the reverse to be true. Rubin made a post not too long ago about different feature reduction methods and we found those to do better than PCA as well. </p>",
      "rawMarkdown": "Commenting on this in case anyone else is working. Our team is using a neural net and when fitting/validating just on the training data less features do better than more features. However, when we fitted to the whole training set and submitted to kaggle, we actually found the reverse to be true. Rubin made a post not too long ago about different feature reduction methods and we found those to do better than PCA as well.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1617294,
      "author_name": "petergribizis",
      "author_url": "",
      "post_date": "12/14/2021 02:32:11",
      "content": "<p>Commenting on this in case anyone else is working. Our team is using a neural net and when fitting/validating just on the training data less features do better than more features. However, when we fitted to the whole training set and submitted to kaggle, we actually found the reverse to be true. Rubin made a post not too long ago about different feature reduction methods and we found those to do better than PCA as well. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1615275": "Adding on to [Aarushi's post](https://www.kaggle.com/c/cs5785-fall-2021-final/discussion/293537), I did PCA for n_components = 50 - 200 and ran a ridge regression on each of them. I found out that lower number of features yields better results for validation. \n\n![](https://i.postimg.cc/bNyvpLnP/Screen-Shot-2021-12-11-at-11-51-59-PM.png)\n\nHope this is helpful! 😊",
    "1617294": "Commenting on this in case anyone else is working. Our team is using a neural net and when fitting/validating just on the training data less features do better than more features. However, when we fitted to the whole training set and submitted to kaggle, we actually found the reverse to be true. Rubin made a post not too long ago about different feature reduction methods and we found those to do better than PCA as well."
  },
  "source": "meta"
}