{
  "id": 68288,
  "title": "Top R Solutions?",
  "url": "/competitions/ga-customer-revenue-prediction/discussion/68288",
  "author_name": "",
  "post_date": "2018-10-11T05:40:45.984583800Z",
  "votes": 2,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi everybody,\nI'm wondering -</p>\n\n<p>Any of you in top ranking (top 50) in LB have a main solution in R?</p>",
  "messages": [
    {
      "id": "402052",
      "postDate": "10/11/2018 05:40:45",
      "content": "<p>Hi everybody,\nI'm wondering -</p>\n\n<p>Any of you in top ranking (top 50) in LB have a main solution in R?</p>",
      "rawMarkdown": "Hi everybody,\nI'm wondering -\n\nAny of you in top ranking (top 50) in LB have a main solution in R?",
      "votes": null
    },
    {
      "id": "402266",
      "postDate": "10/11/2018 12:39:47",
      "content": "<p>AmirH, even though I'm not in top 50, I would like to share some info. I hope this does not pollute your topic.</p>\n\n<p>LB: 1.4114 - Ensemble: R (lightgbm + xgboost) + python (best public kernel)\nLB: 1.4173 - Ensemble: R (lightgbm + xgboost)</p>\n\n<p>I hope more models can add diversity enough to improve my only-R solution. And I have a lot of FE work to do.</p>",
      "rawMarkdown": "AmirH, even though I'm not in top 50, I would like to share some info. I hope this does not pollute your topic.\n\nLB: 1.4114 - Ensemble: R (lightgbm + xgboost) + python (best public kernel)\nLB: 1.4173 - Ensemble: R (lightgbm + xgboost)\n\nI hope more models can add diversity enough to improve my only-R solution. And I have a lot of FE work to do.",
      "votes": null
    },
    {
      "id": "402280",
      "postDate": "10/11/2018 13:01:28",
      "content": "<p>I have replicated the two-stage approach from Oivier (<a href=\"https://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions\">https://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions</a>) in R. In fact have a kernel for that but it is not yet public as there is still a subtle error in it  that I do not have locally.  With a few good data sets you can construct miriads of predictions for an ensemble, because you can feed the preds from dataset1 + model a =&gt; (aggregated) dataset2 + model b .  (So with 3 datasets and 3 models you can build 81 of them).  </p>\n\n<p>II will publish the kernel when it works, but given the current state of this competition I do not know how to continue after that. </p>",
      "rawMarkdown": "I have replicated the two-stage approach from Oivier (https://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions) in R. In fact have a kernel for that but it is not yet public as there is still a subtle error in it  that I do not have locally.  With a few good data sets you can construct miriads of predictions for an ensemble, because you can feed the preds from dataset1 + model a =&gt; (aggregated) dataset2 + model b .  (So with 3 datasets and 3 models you can build 81 of them).  \n\nII will publish the kernel when it works, but given the current state of this competition I do not know how to continue after that.",
      "votes": null
    },
    {
      "id": "402381",
      "postDate": "10/11/2018 15:59:20",
      "content": "<p>Using Olivier's approach I could get LB 1.4072 with single lgbm but without \"pageview / hit ratio\".</p>",
      "rawMarkdown": "Using Olivier's approach I could get LB 1.4072 with single lgbm but without \"pageview / hit ratio\".",
      "votes": null
    },
    {
      "id": "402399",
      "postDate": "10/11/2018 16:31:20",
      "content": "<p>Are you using any external data? thanks!</p>",
      "rawMarkdown": "Are you using any external data? thanks!",
      "votes": null
    },
    {
      "id": "402401",
      "postDate": "10/11/2018 16:33:53",
      "content": "<p>Nope.</p>",
      "rawMarkdown": "Nope.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 402266,
      "author_name": "paulofelipe",
      "author_url": "",
      "post_date": "10/11/2018 12:39:47",
      "content": "<p>AmirH, even though I'm not in top 50, I would like to share some info. I hope this does not pollute your topic.</p>\n\n<p>LB: 1.4114 - Ensemble: R (lightgbm + xgboost) + python (best public kernel)\nLB: 1.4173 - Ensemble: R (lightgbm + xgboost)</p>\n\n<p>I hope more models can add diversity enough to improve my only-R solution. And I have a lot of FE work to do.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 402280,
      "author_name": "mrooijer",
      "author_url": "",
      "post_date": "10/11/2018 13:01:28",
      "content": "<p>I have replicated the two-stage approach from Oivier (<a href=\"https://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions\">https://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions</a>) in R. In fact have a kernel for that but it is not yet public as there is still a subtle error in it  that I do not have locally.  With a few good data sets you can construct miriads of predictions for an ensemble, because you can feed the preds from dataset1 + model a =&gt; (aggregated) dataset2 + model b .  (So with 3 datasets and 3 models you can build 81 of them).  </p>\n\n<p>II will publish the kernel when it works, but given the current state of this competition I do not know how to continue after that. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 402381,
      "author_name": "andrew60909",
      "author_url": "",
      "post_date": "10/11/2018 15:59:20",
      "content": "<p>Using Olivier's approach I could get LB 1.4072 with single lgbm but without \"pageview / hit ratio\".</p>",
      "votes": null,
      "replies": [
        {
          "id": 402399,
          "author_name": "paulofelipe",
          "author_url": "",
          "post_date": "10/11/2018 16:31:20",
          "content": "<p>Are you using any external data? thanks!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 402401,
          "author_name": "andrew60909",
          "author_url": "",
          "post_date": "10/11/2018 16:33:53",
          "content": "<p>Nope.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "402052": "Hi everybody,\nI'm wondering -\n\nAny of you in top ranking (top 50) in LB have a main solution in R?",
    "402266": "AmirH, even though I'm not in top 50, I would like to share some info. I hope this does not pollute your topic.\n\nLB: 1.4114 - Ensemble: R (lightgbm + xgboost) + python (best public kernel)\nLB: 1.4173 - Ensemble: R (lightgbm + xgboost)\n\nI hope more models can add diversity enough to improve my only-R solution. And I have a lot of FE work to do.",
    "402280": "I have replicated the two-stage approach from Oivier (https://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions) in R. In fact have a kernel for that but it is not yet public as there is still a subtle error in it  that I do not have locally.  With a few good data sets you can construct miriads of predictions for an ensemble, because you can feed the preds from dataset1 + model a =&gt; (aggregated) dataset2 + model b .  (So with 3 datasets and 3 models you can build 81 of them).  \n\nII will publish the kernel when it works, but given the current state of this competition I do not know how to continue after that.",
    "402381": "Using Olivier's approach I could get LB 1.4072 with single lgbm but without \"pageview / hit ratio\".",
    "402399": "Are you using any external data? thanks!",
    "402401": "Nope."
  },
  "source": "meta"
}