{
  "id": 53752,
  "title": "Feature Engineering Doubt",
  "url": "/competitions/talkingdata-adtracking-fraud-detection/discussion/53752",
  "author_name": "",
  "post_date": "2018-04-04T17:06:30.381967500Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>First Kaggle Competition..\nI have read almost all discussion</p>\n\n<p>Since it is known that with good feature engineering and single model we can easily cross auc 0.97\nNote ** Good Validation also contributed in crossing 0.97 auc</p>\n\n<p>My mentor said that Good Feature Engineering  is an ART\ni know that it takes lot of practice and learning \ntill now\napart from couting ips and groupby i  m not able to come up with new feature \nso, what i want to ask </p>\n\n<p><strong>How to <em>approach</em> for feature Engineering  ?</strong></p>\n\n<p>1) Hit and try</p>\n\n<p>2) By EDA </p>\n\n<p>3) Domain knowlege</p>\n\n<p>How EDA is important in building new feature ???</p>",
  "messages": [
    {
      "id": "309114",
      "postDate": "04/04/2018 17:06:30",
      "content": "<p>First Kaggle Competition..\nI have read almost all discussion</p>\n\n<p>Since it is known that with good feature engineering and single model we can easily cross auc 0.97\nNote ** Good Validation also contributed in crossing 0.97 auc</p>\n\n<p>My mentor said that Good Feature Engineering  is an ART\ni know that it takes lot of practice and learning \ntill now\napart from couting ips and groupby i  m not able to come up with new feature \nso, what i want to ask </p>\n\n<p><strong>How to <em>approach</em> for feature Engineering  ?</strong></p>\n\n<p>1) Hit and try</p>\n\n<p>2) By EDA </p>\n\n<p>3) Domain knowlege</p>\n\n<p>How EDA is important in building new feature ???</p>",
      "rawMarkdown": "First Kaggle Competition..\nI have read almost all discussion\n\nSince it is known that with good feature engineering and single model we can easily cross auc 0.97\nNote ** Good Validation also contributed in crossing 0.97 auc\n\n\nMy mentor said that Good Feature Engineering  is an ART\ni know that it takes lot of practice and learning \ntill now\napart from couting ips and groupby i  m not able to come up with new feature \nso, what i want to ask \n\n\n**How to *approach* for feature Engineering  ?**\n\n1) Hit and try\n\n2) By EDA \n\n3) Domain knowlege\n\n\nHow EDA is important in building new feature ???",
      "votes": null
    },
    {
      "id": "309124",
      "postDate": "04/04/2018 17:23:51",
      "content": "<p>Good feature engineering is to know your learning algorithm's weaknesses. Feature engineering is basically to make the problem understandable by your model. While creating a feature as a linear combination of others is meaningless in neural nets, it may help your tree based models.</p>",
      "rawMarkdown": "Good feature engineering is to know your learning algorithm's weaknesses. Feature engineering is basically to make the problem understandable by your model. While creating a feature as a linear combination of others is meaningless in neural nets, it may help your tree based models.",
      "votes": null
    },
    {
      "id": "309127",
      "postDate": "04/04/2018 17:27:22",
      "content": "<p>for xgboost can we find model weakness </p>",
      "rawMarkdown": "for xgboost can we find model weakness",
      "votes": null
    },
    {
      "id": "309137",
      "postDate": "04/04/2018 17:45:25",
      "content": "<p>You can try to teach simple summation or multiplication operation with the numbers below (let's say) 1000. And then test with the numbers larger than 1000. Trees have difficulty to extrapolate. </p>",
      "rawMarkdown": "You can try to teach simple summation or multiplication operation with the numbers below (let's say) 1000. And then test with the numbers larger than 1000. Trees have difficulty to extrapolate.",
      "votes": null
    },
    {
      "id": "309313",
      "postDate": "04/05/2018 03:01:55",
      "content": "<p>thanks for intution</p>",
      "rawMarkdown": "thanks for intution",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 309124,
      "author_name": "aerdem4",
      "author_url": "",
      "post_date": "04/04/2018 17:23:51",
      "content": "<p>Good feature engineering is to know your learning algorithm's weaknesses. Feature engineering is basically to make the problem understandable by your model. While creating a feature as a linear combination of others is meaningless in neural nets, it may help your tree based models.</p>",
      "votes": null,
      "replies": [
        {
          "id": 309127,
          "author_name": "shubham95pandey",
          "author_url": "",
          "post_date": "04/04/2018 17:27:22",
          "content": "<p>for xgboost can we find model weakness </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 309137,
          "author_name": "aerdem4",
          "author_url": "",
          "post_date": "04/04/2018 17:45:25",
          "content": "<p>You can try to teach simple summation or multiplication operation with the numbers below (let's say) 1000. And then test with the numbers larger than 1000. Trees have difficulty to extrapolate. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 309313,
          "author_name": "shubham95pandey",
          "author_url": "",
          "post_date": "04/05/2018 03:01:55",
          "content": "<p>thanks for intution</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "309114": "First Kaggle Competition..\nI have read almost all discussion\n\nSince it is known that with good feature engineering and single model we can easily cross auc 0.97\nNote ** Good Validation also contributed in crossing 0.97 auc\n\n\nMy mentor said that Good Feature Engineering  is an ART\ni know that it takes lot of practice and learning \ntill now\napart from couting ips and groupby i  m not able to come up with new feature \nso, what i want to ask \n\n\n**How to *approach* for feature Engineering  ?**\n\n1) Hit and try\n\n2) By EDA \n\n3) Domain knowlege\n\n\nHow EDA is important in building new feature ???",
    "309124": "Good feature engineering is to know your learning algorithm's weaknesses. Feature engineering is basically to make the problem understandable by your model. While creating a feature as a linear combination of others is meaningless in neural nets, it may help your tree based models.",
    "309127": "for xgboost can we find model weakness",
    "309137": "You can try to teach simple summation or multiplication operation with the numbers below (let's say) 1000. And then test with the numbers larger than 1000. Trees have difficulty to extrapolate.",
    "309313": "thanks for intution"
  },
  "source": "meta"
}