{
  "id": 78576,
  "title": "Is it worthwhile to first run a classification algorithm?",
  "url": "/competitions/elo-merchant-category-recommendation/discussion/78576",
  "author_name": "",
  "post_date": "2019-01-25T14:34:49.011891700Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "<p>There is a bunch of outliers in the training dataset with the same 'target' value. Thus, one might think that there could be two classes of customers. Ones with the outlier target value, and the others whose target value is on a continuous scale...</p>\n\n<p>Would it be worthwhile strategy to first try to find a way to classify customers to these 2 groups and then apply the regression algorithm only to those predicted as 'non-outliers'?</p>",
  "messages": [
    {
      "id": "461209",
      "postDate": "01/25/2019 14:34:49",
      "content": "<p>There is a bunch of outliers in the training dataset with the same 'target' value. Thus, one might think that there could be two classes of customers. Ones with the outlier target value, and the others whose target value is on a continuous scale...</p>\n\n<p>Would it be worthwhile strategy to first try to find a way to classify customers to these 2 groups and then apply the regression algorithm only to those predicted as 'non-outliers'?</p>",
      "rawMarkdown": "There is a bunch of outliers in the training dataset with the same 'target' value. Thus, one might think that there could be two classes of customers. Ones with the outlier target value, and the others whose target value is on a continuous scale...\n\nWould it be worthwhile strategy to first try to find a way to classify customers to these 2 groups and then apply the regression algorithm only to those predicted as 'non-outliers'?",
      "votes": null
    },
    {
      "id": "461287",
      "postDate": "01/25/2019 17:45:19",
      "content": "<p>It could be worth a look, but in terms of return-on-time your best bet is likely to go straight for bounded regression from the start. With the same input variables, depending on the models used, the use of an \"outlier indicator\" as an input for a subsequent regression model would likely not add much value and may increase your chance of overfitting. If there was a different sort of classification that you could perform initially, that could still have value, but specifically whether or not they're an outlier will likely not be a great input variable unless your final model is highly naive. </p>",
      "rawMarkdown": "It could be worth a look, but in terms of return-on-time your best bet is likely to go straight for bounded regression from the start. With the same input variables, depending on the models used, the use of an \"outlier indicator\" as an input for a subsequent regression model would likely not add much value and may increase your chance of overfitting. If there was a different sort of classification that you could perform initially, that could still have value, but specifically whether or not they're an outlier will likely not be a great input variable unless your final model is highly naive.",
      "votes": null
    },
    {
      "id": "461705",
      "postDate": "01/26/2019 19:35:44",
      "content": "<p>It is for sure. Even a lot of people in this competition are using that way. \nsee this : <a href=\"https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/75173\">https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/75173</a>. In the end it boils down to building the best outlier classifier. </p>",
      "rawMarkdown": "It is for sure. Even a lot of people in this competition are using that way. \nsee this : https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/75173. In the end it boils down to building the best outlier classifier.",
      "votes": null
    },
    {
      "id": "462096",
      "postDate": "01/27/2019 17:02:54",
      "content": "<p>Thank you very much for you response</p>",
      "rawMarkdown": "Thank you very much for you response",
      "votes": null
    },
    {
      "id": "462097",
      "postDate": "01/27/2019 17:03:02",
      "content": "<p>Thank you!</p>",
      "rawMarkdown": "Thank you!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 461287,
      "author_name": "alecthekulak",
      "author_url": "",
      "post_date": "01/25/2019 17:45:19",
      "content": "<p>It could be worth a look, but in terms of return-on-time your best bet is likely to go straight for bounded regression from the start. With the same input variables, depending on the models used, the use of an \"outlier indicator\" as an input for a subsequent regression model would likely not add much value and may increase your chance of overfitting. If there was a different sort of classification that you could perform initially, that could still have value, but specifically whether or not they're an outlier will likely not be a great input variable unless your final model is highly naive. </p>",
      "votes": null,
      "replies": [
        {
          "id": 462097,
          "author_name": "saki24",
          "author_url": "",
          "post_date": "01/27/2019 17:03:02",
          "content": "<p>Thank you!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 461705,
      "author_name": "harshit92",
      "author_url": "",
      "post_date": "01/26/2019 19:35:44",
      "content": "<p>It is for sure. Even a lot of people in this competition are using that way. \nsee this : <a href=\"https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/75173\">https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/75173</a>. In the end it boils down to building the best outlier classifier. </p>",
      "votes": null,
      "replies": [
        {
          "id": 462096,
          "author_name": "saki24",
          "author_url": "",
          "post_date": "01/27/2019 17:02:54",
          "content": "<p>Thank you very much for you response</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "461209": "There is a bunch of outliers in the training dataset with the same 'target' value. Thus, one might think that there could be two classes of customers. Ones with the outlier target value, and the others whose target value is on a continuous scale...\n\nWould it be worthwhile strategy to first try to find a way to classify customers to these 2 groups and then apply the regression algorithm only to those predicted as 'non-outliers'?",
    "461287": "It could be worth a look, but in terms of return-on-time your best bet is likely to go straight for bounded regression from the start. With the same input variables, depending on the models used, the use of an \"outlier indicator\" as an input for a subsequent regression model would likely not add much value and may increase your chance of overfitting. If there was a different sort of classification that you could perform initially, that could still have value, but specifically whether or not they're an outlier will likely not be a great input variable unless your final model is highly naive.",
    "461705": "It is for sure. Even a lot of people in this competition are using that way. \nsee this : https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/75173. In the end it boils down to building the best outlier classifier.",
    "462096": "Thank you very much for you response",
    "462097": "Thank you!"
  },
  "source": "meta"
}