{
  "id": 294169,
  "title": "PCA on ResNet Features",
  "url": "/competitions/cs5785-fall-2021-final/discussion/294169",
  "author_name": "rfaures",
  "post_date": "2021-12-08T20:30:52.993000",
  "votes": 10,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Building on <a href=\"https://www.kaggle.com/c/cs5785-fall-2021-final/discussion/293537\" target=\"_blank\">Aarushi's post</a> on PCA feature selection.</p>\n<p>We wanted to apply PCA to reduce the dimensionality of the ResNet features. It had a significant positive impact on our MAPE. Looking at the cumulative variance explained by the principal components, we see that the 100 first components explain 94% of the variance, and the 200 first components explain 99% of the variance.</p>\n<p><a href=\"https://ibb.co/C600dJQ\"><img src=\"https://i.ibb.co/bFddhHX/t-l-chargement.png\" alt=\"t-l-chargement\"></a></p>\n<p>We are looking for a good trade-off between reaching lower dimensions without losing too much information.</p>",
  "messages": [
    {
      "id": 1612374,
      "postDate": "2021-12-08T20:30:52.993Z",
      "content": "<p>Building on <a href=\"https://www.kaggle.com/c/cs5785-fall-2021-final/discussion/293537\" target=\"_blank\">Aarushi's post</a> on PCA feature selection.</p>\n<p>We wanted to apply PCA to reduce the dimensionality of the ResNet features. It had a significant positive impact on our MAPE. Looking at the cumulative variance explained by the principal components, we see that the 100 first components explain 94% of the variance, and the 200 first components explain 99% of the variance.</p>\n<p><a href=\"https://ibb.co/C600dJQ\"><img src=\"https://i.ibb.co/bFddhHX/t-l-chargement.png\" alt=\"t-l-chargement\"></a></p>\n<p>We are looking for a good trade-off between reaching lower dimensions without losing too much information.</p>",
      "rawMarkdown": "Building on [Aarushi's post](https://www.kaggle.com/c/cs5785-fall-2021-final/discussion/293537) on PCA feature selection.\n\nWe wanted to apply PCA to reduce the dimensionality of the ResNet features. It had a significant positive impact on our MAPE. Looking at the cumulative variance explained by the principal components, we see that the 100 first components explain 94% of the variance, and the 200 first components explain 99% of the variance.\n\n<a href=\"https://ibb.co/C600dJQ\"><img src=\"https://i.ibb.co/bFddhHX/t-l-chargement.png\" alt=\"t-l-chargement\" border=\"0\"></a>\n\nWe are looking for a good trade-off between reaching lower dimensions without losing too much information.",
      "votes": 9
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1612374": "Building on [Aarushi's post](https://www.kaggle.com/c/cs5785-fall-2021-final/discussion/293537) on PCA feature selection.\n\nWe wanted to apply PCA to reduce the dimensionality of the ResNet features. It had a significant positive impact on our MAPE. Looking at the cumulative variance explained by the principal components, we see that the 100 first components explain 94% of the variance, and the 200 first components explain 99% of the variance.\n\n<a href=\"https://ibb.co/C600dJQ\"><img src=\"https://i.ibb.co/bFddhHX/t-l-chargement.png\" alt=\"t-l-chargement\" border=\"0\"></a>\n\nWe are looking for a good trade-off between reaching lower dimensions without losing too much information."
  }
}