{
  "id": 125998,
  "title": "Next Steps after Best Performing Models/Variable Importance",
  "url": "/competitions/cat-in-the-dat-ii/discussion/125998",
  "author_name": "",
  "post_date": "2020-01-15T02:35:28.518765Z",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi D.S. community,</p>\n\n<p>I really love the technique from @ganeshmundra of running a handful of models through a dataframe and printing out the variable importance.</p>\n\n<p>https://www.kaggle.com/ganeshmundra/prediction-using-6-diff-model</p>\n\n<p>After this step, it seems like the take aways are which type of model appears to be the best and the most important variables. However, each model gives completely different variable importance.</p>\n\n<p>After this great analysis, what is the next logical step? How can we use the information from running a handful of models for improvement. I assume this would lead to focusing on certain features for feature engineering.</p>",
  "messages": [
    {
      "id": "719004",
      "postDate": "01/15/2020 02:35:28",
      "content": "<p>Hi D.S. community,</p>\n\n<p>I really love the technique from @ganeshmundra of running a handful of models through a dataframe and printing out the variable importance.</p>\n\n<p>https://www.kaggle.com/ganeshmundra/prediction-using-6-diff-model</p>\n\n<p>After this step, it seems like the take aways are which type of model appears to be the best and the most important variables. However, each model gives completely different variable importance.</p>\n\n<p>After this great analysis, what is the next logical step? How can we use the information from running a handful of models for improvement. I assume this would lead to focusing on certain features for feature engineering.</p>",
      "rawMarkdown": "Hi D.S. community,\n\nI really love the technique from @ganeshmundra of running a handful of models through a dataframe and printing out the variable importance.\n\n[https://www.kaggle.com/ganeshmundra/prediction-using-6-diff-model]()\n\nAfter this step, it seems like the take aways are which type of model appears to be the best and the most important variables. However, each model gives completely different variable importance.\n\nAfter this great analysis, what is the next logical step? How can we use the information from running a handful of models for improvement. I assume this would lead to focusing on certain features for feature engineering.",
      "votes": null
    },
    {
      "id": "719096",
      "postDate": "01/15/2020 06:01:39",
      "content": "<p>Hello. In the strong classifiers it seems like high-cardinality nominal variables are more important than the others. They are generally in top 5. So you have to work on them primarily in my opinion. Secondly, filling missing values in all variables with mode is risky. It also affects feature importances. But it won't affect high-card variables too much. So maybe you can group your dataframe by them and fill with similar values.</p>\n\n<p>In addition, there are several ways to handle categorical features. For example check here and you will see 15 different kind of encoding techniques: <a href=\"https://contrib.scikit-learn.org/categorical-encoding/index.html\">https://contrib.scikit-learn.org/categorical-encoding/index.html</a></p>",
      "rawMarkdown": "Hello. In the strong classifiers it seems like high-cardinality nominal variables are more important than the others. They are generally in top 5. So you have to work on them primarily in my opinion. Secondly, filling missing values in all variables with mode is risky. It also affects feature importances. But it won't affect high-card variables too much. So maybe you can group your dataframe by them and fill with similar values.\n\nIn addition, there are several ways to handle categorical features. For example check here and you will see 15 different kind of encoding techniques: https://contrib.scikit-learn.org/categorical-encoding/index.html",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 719096,
      "author_name": "cemsarier",
      "author_url": "",
      "post_date": "01/15/2020 06:01:39",
      "content": "<p>Hello. In the strong classifiers it seems like high-cardinality nominal variables are more important than the others. They are generally in top 5. So you have to work on them primarily in my opinion. Secondly, filling missing values in all variables with mode is risky. It also affects feature importances. But it won't affect high-card variables too much. So maybe you can group your dataframe by them and fill with similar values.</p>\n\n<p>In addition, there are several ways to handle categorical features. For example check here and you will see 15 different kind of encoding techniques: <a href=\"https://contrib.scikit-learn.org/categorical-encoding/index.html\">https://contrib.scikit-learn.org/categorical-encoding/index.html</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "719004": "Hi D.S. community,\n\nI really love the technique from @ganeshmundra of running a handful of models through a dataframe and printing out the variable importance.\n\n[https://www.kaggle.com/ganeshmundra/prediction-using-6-diff-model]()\n\nAfter this step, it seems like the take aways are which type of model appears to be the best and the most important variables. However, each model gives completely different variable importance.\n\nAfter this great analysis, what is the next logical step? How can we use the information from running a handful of models for improvement. I assume this would lead to focusing on certain features for feature engineering.",
    "719096": "Hello. In the strong classifiers it seems like high-cardinality nominal variables are more important than the others. They are generally in top 5. So you have to work on them primarily in my opinion. Secondly, filling missing values in all variables with mode is risky. It also affects feature importances. But it won't affect high-card variables too much. So maybe you can group your dataframe by them and fill with similar values.\n\nIn addition, there are several ways to handle categorical features. For example check here and you will see 15 different kind of encoding techniques: https://contrib.scikit-learn.org/categorical-encoding/index.html"
  },
  "source": "meta"
}