{
  "id": 72075,
  "title": "Can you share your ideas about how to make a model",
  "url": "/competitions/ga-customer-revenue-prediction/discussion/72075",
  "author_name": "",
  "post_date": "2018-11-20T04:56:39.536969300Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>In this competition we have to predict for the future and we don't have features in future. So how are you predicting the target ?</p>",
  "messages": [
    {
      "id": "424415",
      "postDate": "11/20/2018 04:56:39",
      "content": "<p>In this competition we have to predict for the future and we don't have features in future. So how are you predicting the target ?</p>",
      "rawMarkdown": "In this competition we have to predict for the future and we don't have features in future. So how are you predicting the target ?",
      "votes": null
    },
    {
      "id": "424545",
      "postDate": "11/20/2018 09:48:29",
      "content": "<p>Hello Suchith,</p>\n\n<p>For this competition this is the model i tried : KNN -&gt; Classification -&gt; Regression. </p>\n\n<p>Basically, I used KNN to create groups of similar users to, first classify if they will spend something, then, if they are classified as spending something, i use regression to know \"how much\" they will spend.</p>",
      "rawMarkdown": "Hello Suchith,\n\nFor this competition this is the model i tried : KNN -&gt; Classification -&gt; Regression. \n\nBasically, I used KNN to create groups of similar users to, first classify if they will spend something, then, if they are classified as spending something, i use regression to know \"how much\" they will spend.",
      "votes": null
    },
    {
      "id": "424916",
      "postDate": "11/20/2018 21:56:42",
      "content": "<p>This is a time series prediction problem, using only time-related features since buyers will be the same and therefore user-related variables may consider them as constants in time. I recommend using Recurrent Neural Networks</p>",
      "rawMarkdown": "This is a time series prediction problem, using only time-related features since buyers will be the same and therefore user-related variables may consider them as constants in time. I recommend using Recurrent Neural Networks",
      "votes": null
    },
    {
      "id": "427220",
      "postDate": "11/24/2018 21:50:31",
      "content": "<p>Try to think broadly of the time periods involved. </p>\n\n<p>We are predicting the 62 days from 2018-12-01 to 2019-01-31 with historic data ending at 2018-10-15. The difference between the first day we are predicting and the last day of historic data is 48 days (i.e. 2018-12-01 minus 2018-10-15). So one, way to think of the problem is predicting a 62 day interval 48 days ahead of now. </p>\n\n<p>Given the above formulation, you can then create training and test sets based purely on the historic data. In other words, 2018-08-15 to 2018-10-15 (the last 62 days of the historic data) could represent your test set, with all of the data before 2018-06-29 being your training data. </p>\n\n<p>From that training data (i.e. everything before 2018-06-29) you could then create general features such as \"spend_in_the_last_30_days\" or \"number_of_pg_visits_in_last_90_days\" to predict the interval of 2018-08-15 to 2018-10-15. Since these features are generalized you can then use them to predict the future period of 2018-12-01 to 2019-01-31. </p>\n\n<p>The challenge is trying to think of more of these general types of features that might be available. </p>",
      "rawMarkdown": "Try to think broadly of the time periods involved. \n\nWe are predicting the 62 days from 2018-12-01 to 2019-01-31 with historic data ending at 2018-10-15. The difference between the first day we are predicting and the last day of historic data is 48 days (i.e. 2018-12-01 minus 2018-10-15). So one, way to think of the problem is predicting a 62 day interval 48 days ahead of now. \n\nGiven the above formulation, you can then create training and test sets based purely on the historic data. In other words, 2018-08-15 to 2018-10-15 (the last 62 days of the historic data) could represent your test set, with all of the data before 2018-06-29 being your training data. \n\nFrom that training data (i.e. everything before 2018-06-29) you could then create general features such as \"spend_in_the_last_30_days\" or \"number_of_pg_visits_in_last_90_days\" to predict the interval of 2018-08-15 to 2018-10-15. Since these features are generalized you can then use them to predict the future period of 2018-12-01 to 2019-01-31. \n\nThe challenge is trying to think of more of these general types of features that might be available.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 424545,
      "author_name": "xtelima",
      "author_url": "",
      "post_date": "11/20/2018 09:48:29",
      "content": "<p>Hello Suchith,</p>\n\n<p>For this competition this is the model i tried : KNN -&gt; Classification -&gt; Regression. </p>\n\n<p>Basically, I used KNN to create groups of similar users to, first classify if they will spend something, then, if they are classified as spending something, i use regression to know \"how much\" they will spend.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 424916,
      "author_name": "paulorzp",
      "author_url": "",
      "post_date": "11/20/2018 21:56:42",
      "content": "<p>This is a time series prediction problem, using only time-related features since buyers will be the same and therefore user-related variables may consider them as constants in time. I recommend using Recurrent Neural Networks</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 427220,
      "author_name": "johepaton",
      "author_url": "",
      "post_date": "11/24/2018 21:50:31",
      "content": "<p>Try to think broadly of the time periods involved. </p>\n\n<p>We are predicting the 62 days from 2018-12-01 to 2019-01-31 with historic data ending at 2018-10-15. The difference between the first day we are predicting and the last day of historic data is 48 days (i.e. 2018-12-01 minus 2018-10-15). So one, way to think of the problem is predicting a 62 day interval 48 days ahead of now. </p>\n\n<p>Given the above formulation, you can then create training and test sets based purely on the historic data. In other words, 2018-08-15 to 2018-10-15 (the last 62 days of the historic data) could represent your test set, with all of the data before 2018-06-29 being your training data. </p>\n\n<p>From that training data (i.e. everything before 2018-06-29) you could then create general features such as \"spend_in_the_last_30_days\" or \"number_of_pg_visits_in_last_90_days\" to predict the interval of 2018-08-15 to 2018-10-15. Since these features are generalized you can then use them to predict the future period of 2018-12-01 to 2019-01-31. </p>\n\n<p>The challenge is trying to think of more of these general types of features that might be available. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "424415": "In this competition we have to predict for the future and we don't have features in future. So how are you predicting the target ?",
    "424545": "Hello Suchith,\n\nFor this competition this is the model i tried : KNN -&gt; Classification -&gt; Regression. \n\nBasically, I used KNN to create groups of similar users to, first classify if they will spend something, then, if they are classified as spending something, i use regression to know \"how much\" they will spend.",
    "424916": "This is a time series prediction problem, using only time-related features since buyers will be the same and therefore user-related variables may consider them as constants in time. I recommend using Recurrent Neural Networks",
    "427220": "Try to think broadly of the time periods involved. \n\nWe are predicting the 62 days from 2018-12-01 to 2019-01-31 with historic data ending at 2018-10-15. The difference between the first day we are predicting and the last day of historic data is 48 days (i.e. 2018-12-01 minus 2018-10-15). So one, way to think of the problem is predicting a 62 day interval 48 days ahead of now. \n\nGiven the above formulation, you can then create training and test sets based purely on the historic data. In other words, 2018-08-15 to 2018-10-15 (the last 62 days of the historic data) could represent your test set, with all of the data before 2018-06-29 being your training data. \n\nFrom that training data (i.e. everything before 2018-06-29) you could then create general features such as \"spend_in_the_last_30_days\" or \"number_of_pg_visits_in_last_90_days\" to predict the interval of 2018-08-15 to 2018-10-15. Since these features are generalized you can then use them to predict the future period of 2018-12-01 to 2019-01-31. \n\nThe challenge is trying to think of more of these general types of features that might be available."
  },
  "source": "meta"
}