{
  "id": 126862,
  "title": "Fancy imputation?",
  "url": "/competitions/cat-in-the-dat-ii/discussion/126862",
  "author_name": "",
  "post_date": "2020-01-20T19:55:35.422743900Z",
  "votes": 8,
  "comment_count": 4,
  "views": 0,
  "content": "<p>As someone else mentioned: There's something interesting about the missing values. </p>\n\n<p>So far I've avoided any imputation at all, maintaining the missing values as their own class as much as possible. I've noticed that some others are using <strong>simple imputation</strong> via mode or median, but I haven't found that to approach to be particularly useful myself. I think maybe it's so simple that it results in a loss of information.</p>\n\n<p>But I wonder if a more complex imputation strategy might be useful, and if understanding variable interactions might be key to running a complex / multivariate imputation? Has anyone explored <strong>fancy imputation</strong>?</p>\n\n<p>I'm thinking this package might be useful:</p>\n\n<p><a href=\"https://github.com/iskandr/fancyimpute\">https://github.com/iskandr/fancyimpute</a></p>\n\n<p>Or scikit's IterativeImputer:</p>\n\n<p><a href=\"https://scikit-learn.org/stable/modules/generated/sklearn.impute.IterativeImputer.html\">https://scikit-learn.org/stable/modules/generated/sklearn.impute.IterativeImputer.html</a></p>",
  "messages": [
    {
      "id": "724103",
      "postDate": "01/20/2020 19:55:35",
      "content": "<p>As someone else mentioned: There's something interesting about the missing values. </p>\n\n<p>So far I've avoided any imputation at all, maintaining the missing values as their own class as much as possible. I've noticed that some others are using <strong>simple imputation</strong> via mode or median, but I haven't found that to approach to be particularly useful myself. I think maybe it's so simple that it results in a loss of information.</p>\n\n<p>But I wonder if a more complex imputation strategy might be useful, and if understanding variable interactions might be key to running a complex / multivariate imputation? Has anyone explored <strong>fancy imputation</strong>?</p>\n\n<p>I'm thinking this package might be useful:</p>\n\n<p><a href=\"https://github.com/iskandr/fancyimpute\">https://github.com/iskandr/fancyimpute</a></p>\n\n<p>Or scikit's IterativeImputer:</p>\n\n<p><a href=\"https://scikit-learn.org/stable/modules/generated/sklearn.impute.IterativeImputer.html\">https://scikit-learn.org/stable/modules/generated/sklearn.impute.IterativeImputer.html</a></p>",
      "rawMarkdown": "As someone else mentioned: There's something interesting about the missing values. \n\nSo far I've avoided any imputation at all, maintaining the missing values as their own class as much as possible. I've noticed that some others are using **simple imputation** via mode or median, but I haven't found that to approach to be particularly useful myself. I think maybe it's so simple that it results in a loss of information.\n\nBut I wonder if a more complex imputation strategy might be useful, and if understanding variable interactions might be key to running a complex / multivariate imputation? Has anyone explored **fancy imputation**?\n\nI'm thinking this package might be useful:\n\nhttps://github.com/iskandr/fancyimpute\n\nOr scikit's IterativeImputer:\n\nhttps://scikit-learn.org/stable/modules/generated/sklearn.impute.IterativeImputer.html",
      "votes": null
    },
    {
      "id": "724169",
      "postDate": "01/20/2020 21:32:45",
      "content": "<p>Since nom_7 was important feature so I grouped on it, and fill missing values. But i am getting low auc.</p>",
      "rawMarkdown": "Since nom_7 was important feature so I grouped on it, and fill missing values. But i am getting low auc.",
      "votes": null
    },
    {
      "id": "729776",
      "postDate": "01/26/2020 16:22:34",
      "content": "<p>I'm thinking about which method for imputation to use, did you tried this fancy imputation <a href=\"/bluewizard\">@bluewizard</a>? How does it worked out?</p>",
      "rawMarkdown": "I'm thinking about which method for imputation to use, did you tried this fancy imputation @bluewizard? How does it worked out?",
      "votes": null
    },
    {
      "id": "730921",
      "postDate": "01/28/2020 05:55:49",
      "content": "<p>Very interesting topic! I haven't jumped into this competition yet, but I came across this on Twitter a while back which might be useful: \n<a href=\"https://github.com/eltonlaw/impyute\">https://github.com/eltonlaw/impyute</a></p>",
      "rawMarkdown": "Very interesting topic! I haven't jumped into this competition yet, but I came across this on Twitter a while back which might be useful: \nhttps://github.com/eltonlaw/impyute",
      "votes": null
    },
    {
      "id": "735713",
      "postDate": "02/03/2020 11:20:27",
      "content": "<p>I have tried to predict missing values with the remaining features except „Target“. Validation via complete cases.</p>",
      "rawMarkdown": "I have tried to predict missing values with the remaining features except „Target“. Validation via complete cases.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 724169,
      "author_name": "shivashi11",
      "author_url": "",
      "post_date": "01/20/2020 21:32:45",
      "content": "<p>Since nom_7 was important feature so I grouped on it, and fill missing values. But i am getting low auc.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 729776,
      "author_name": "brunhs",
      "author_url": "",
      "post_date": "01/26/2020 16:22:34",
      "content": "<p>I'm thinking about which method for imputation to use, did you tried this fancy imputation <a href=\"/bluewizard\">@bluewizard</a>? How does it worked out?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 730921,
      "author_name": "anjum48",
      "author_url": "",
      "post_date": "01/28/2020 05:55:49",
      "content": "<p>Very interesting topic! I haven't jumped into this competition yet, but I came across this on Twitter a while back which might be useful: \n<a href=\"https://github.com/eltonlaw/impyute\">https://github.com/eltonlaw/impyute</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 735713,
      "author_name": "frankmollard",
      "author_url": "",
      "post_date": "02/03/2020 11:20:27",
      "content": "<p>I have tried to predict missing values with the remaining features except „Target“. Validation via complete cases.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "724103": "As someone else mentioned: There's something interesting about the missing values. \n\nSo far I've avoided any imputation at all, maintaining the missing values as their own class as much as possible. I've noticed that some others are using **simple imputation** via mode or median, but I haven't found that to approach to be particularly useful myself. I think maybe it's so simple that it results in a loss of information.\n\nBut I wonder if a more complex imputation strategy might be useful, and if understanding variable interactions might be key to running a complex / multivariate imputation? Has anyone explored **fancy imputation**?\n\nI'm thinking this package might be useful:\n\nhttps://github.com/iskandr/fancyimpute\n\nOr scikit's IterativeImputer:\n\nhttps://scikit-learn.org/stable/modules/generated/sklearn.impute.IterativeImputer.html",
    "724169": "Since nom_7 was important feature so I grouped on it, and fill missing values. But i am getting low auc.",
    "729776": "I'm thinking about which method for imputation to use, did you tried this fancy imputation @bluewizard? How does it worked out?",
    "730921": "Very interesting topic! I haven't jumped into this competition yet, but I came across this on Twitter a while back which might be useful: \nhttps://github.com/eltonlaw/impyute",
    "735713": "I have tried to predict missing values with the remaining features except „Target“. Validation via complete cases."
  },
  "source": "meta"
}