{
  "id": 81328,
  "title": "Small DataSet ?",
  "url": "/competitions/elo-merchant-category-recommendation/discussion/81328",
  "author_name": "",
  "post_date": "2019-02-20T19:38:49.725704300Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Big problem for set is outliers. We have 2207 rows with target = -33.21928095.\nIts rows ~1% of data, but for score  its 55% (zero predict with 2207 correct outliers get 1.7 RMSE).\nHow detect 1% on so small set (2207 examples vs 201917 rows)? Super unbalanced exercise.\nModel can't find pattern for this rows and try use random correlation with features.</p>",
  "messages": [
    {
      "id": "475472",
      "postDate": "02/20/2019 19:38:49",
      "content": "<p>Big problem for set is outliers. We have 2207 rows with target = -33.21928095.\nIts rows ~1% of data, but for score  its 55% (zero predict with 2207 correct outliers get 1.7 RMSE).\nHow detect 1% on so small set (2207 examples vs 201917 rows)? Super unbalanced exercise.\nModel can't find pattern for this rows and try use random correlation with features.</p>",
      "rawMarkdown": "Big problem for set is outliers. We have 2207 rows with target = -33.21928095.\nIts rows ~1% of data, but for score  its 55% (zero predict with 2207 correct outliers get 1.7 RMSE).\nHow detect 1% on so small set (2207 examples vs 201917 rows)? Super unbalanced exercise.\nModel can't find pattern for this rows and try use random correlation with features.",
      "votes": null
    },
    {
      "id": "475549",
      "postDate": "02/20/2019 21:54:46",
      "content": "<p>You could use ROC as the metric to maximize because ROC isn't affected by a class imbalance.  For example, if you predicted all 0 for a card_id being an outlier, you would get a 99% accuracy, but a 0.5 ROC.</p>",
      "rawMarkdown": "You could use ROC as the metric to maximize because ROC isn't affected by a class imbalance.  For example, if you predicted all 0 for a card_id being an outlier, you would get a 99% accuracy, but a 0.5 ROC.",
      "votes": null
    },
    {
      "id": "485018",
      "postDate": "03/06/2019 19:56:35",
      "content": "<p>Hello,\nSee this kernel with some tips to deal with very small datasets: <a href=\"https://www.kaggle.com/rafjaa/dealing-with-very-small-datasets\">https://www.kaggle.com/rafjaa/dealing-with-very-small-datasets</a></p>",
      "rawMarkdown": "Hello,\nSee this kernel with some tips to deal with very small datasets: https://www.kaggle.com/rafjaa/dealing-with-very-small-datasets",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 475549,
      "author_name": "petersk20",
      "author_url": "",
      "post_date": "02/20/2019 21:54:46",
      "content": "<p>You could use ROC as the metric to maximize because ROC isn't affected by a class imbalance.  For example, if you predicted all 0 for a card_id being an outlier, you would get a 99% accuracy, but a 0.5 ROC.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 485018,
      "author_name": "rafjaa",
      "author_url": "",
      "post_date": "03/06/2019 19:56:35",
      "content": "<p>Hello,\nSee this kernel with some tips to deal with very small datasets: <a href=\"https://www.kaggle.com/rafjaa/dealing-with-very-small-datasets\">https://www.kaggle.com/rafjaa/dealing-with-very-small-datasets</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "475472": "Big problem for set is outliers. We have 2207 rows with target = -33.21928095.\nIts rows ~1% of data, but for score  its 55% (zero predict with 2207 correct outliers get 1.7 RMSE).\nHow detect 1% on so small set (2207 examples vs 201917 rows)? Super unbalanced exercise.\nModel can't find pattern for this rows and try use random correlation with features.",
    "475549": "You could use ROC as the metric to maximize because ROC isn't affected by a class imbalance.  For example, if you predicted all 0 for a card_id being an outlier, you would get a 99% accuracy, but a 0.5 ROC.",
    "485018": "Hello,\nSee this kernel with some tips to deal with very small datasets: https://www.kaggle.com/rafjaa/dealing-with-very-small-datasets"
  },
  "source": "meta"
}