{
  "id": 412863,
  "title": "💥 Catboost with meta-model",
  "url": "/competitions/predict-student-performance-from-game-play/discussion/412863",
  "author_name": "",
  "post_date": "2023-05-25T14:34:13.540931600Z",
  "votes": 6,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi there!👋💥</p>\n<p>I created a fresh notebook <a href=\"https://www.kaggle.com/code/ivanisaev/0-697-catboost-model-on-previous-predictions\" target=\"_blank\">https://www.kaggle.com/code/ivanisaev/0-697-catboost-model-on-previous-predictions</a> 🚀</p>\n<p><strong>What is it about:</strong></p>\n<p>📌 In this <a href=\"https://www.kaggle.com/code/ivanisaev/model-based-on-previous-predictions\" target=\"_blank\">notebook</a> I wrote about how to train model on correct labels for other questions in current session. </p>\n<p>📌 For this notebook I similarly trained a meta-XGBoost model but instead of correct labels for training I used predictions (probabilities not labels after converting by threshold) of this model on trained dataset. And I use this meta-model to mix it with main model to obtain final predictions.</p>\n<p><strong>My approach</strong></p>\n<p>📌 As a main model I used model from <a href=\"https://www.kaggle.com/code/vadimkamaev/catboost-mix\" target=\"_blank\">this notebook</a> and trained the meta-model using predictions of this main model on train data.</p>\n<p>📌☝️ In the iter test I save predictions for the previous questions for current and previous level groups. This previous predictions are used in final mix by meta-model to make final predictions more robust.</p>\n<p>📌 To mix predictions of main Catboost-mix model and meta-model I used similar schema as in <a href=\"https://www.kaggle.com/code/ivanisaev/0-697-ensamble-nn-catboost\" target=\"_blank\">my previous notebook</a>. </p>\n<p>📌 Again I didn't receive score improvement in public LB but performance didn't decrease. I hope this is a good sign that model become tore robust for private LB data.</p>\n<p>Will appreciate your feedback! Thank you and have a nice day!🙂</p>",
  "messages": [
    {
      "id": "2273990",
      "postDate": "05/25/2023 14:34:13",
      "content": "<p>Hi there!👋💥</p>\n<p>I created a fresh notebook <a href=\"https://www.kaggle.com/code/ivanisaev/0-697-catboost-model-on-previous-predictions\" target=\"_blank\">https://www.kaggle.com/code/ivanisaev/0-697-catboost-model-on-previous-predictions</a> 🚀</p>\n<p><strong>What is it about:</strong></p>\n<p>📌 In this <a href=\"https://www.kaggle.com/code/ivanisaev/model-based-on-previous-predictions\" target=\"_blank\">notebook</a> I wrote about how to train model on correct labels for other questions in current session. </p>\n<p>📌 For this notebook I similarly trained a meta-XGBoost model but instead of correct labels for training I used predictions (probabilities not labels after converting by threshold) of this model on trained dataset. And I use this meta-model to mix it with main model to obtain final predictions.</p>\n<p><strong>My approach</strong></p>\n<p>📌 As a main model I used model from <a href=\"https://www.kaggle.com/code/vadimkamaev/catboost-mix\" target=\"_blank\">this notebook</a> and trained the meta-model using predictions of this main model on train data.</p>\n<p>📌☝️ In the iter test I save predictions for the previous questions for current and previous level groups. This previous predictions are used in final mix by meta-model to make final predictions more robust.</p>\n<p>📌 To mix predictions of main Catboost-mix model and meta-model I used similar schema as in <a href=\"https://www.kaggle.com/code/ivanisaev/0-697-ensamble-nn-catboost\" target=\"_blank\">my previous notebook</a>. </p>\n<p>📌 Again I didn't receive score improvement in public LB but performance didn't decrease. I hope this is a good sign that model become tore robust for private LB data.</p>\n<p>Will appreciate your feedback! Thank you and have a nice day!🙂</p>",
      "rawMarkdown": "Hi there!👋💥\n\nI created a fresh notebook https://www.kaggle.com/code/ivanisaev/0-697-catboost-model-on-previous-predictions 🚀\n\n**What is it about:**\n\n📌 In this [notebook]( https://www.kaggle.com/code/ivanisaev/model-based-on-previous-predictions) I wrote about how to train model on correct labels for other questions in current session. \n\n📌 For this notebook I similarly trained a meta-XGBoost model but instead of correct labels for training I used predictions (probabilities not labels after converting by threshold) of this model on trained dataset. And I use this meta-model to mix it with main model to obtain final predictions.\n\n**My approach**\n\n📌 As a main model I used model from [this notebook](https://www.kaggle.com/code/vadimkamaev/catboost-mix) and trained the meta-model using predictions of this main model on train data.\n\n📌☝️ In the iter test I save predictions for the previous questions for current and previous level groups. This previous predictions are used in final mix by meta-model to make final predictions more robust.\n \n📌 To mix predictions of main Catboost-mix model and meta-model I used similar schema as in [my previous notebook](https://www.kaggle.com/code/ivanisaev/0-697-ensamble-nn-catboost). \n\n📌 Again I didn't receive score improvement in public LB but performance didn't decrease. I hope this is a good sign that model become tore robust for private LB data.\n\nWill appreciate your feedback! Thank you and have a nice day!🙂",
      "votes": null
    },
    {
      "id": "2292696",
      "postDate": "06/08/2023 15:25:07",
      "content": "<p>Thanks for the notebook.</p>",
      "rawMarkdown": "Thanks for the notebook.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2292696,
      "author_name": "dplg007",
      "author_url": "",
      "post_date": "06/08/2023 15:25:07",
      "content": "<p>Thanks for the notebook.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2273990": "Hi there!👋💥\n\nI created a fresh notebook https://www.kaggle.com/code/ivanisaev/0-697-catboost-model-on-previous-predictions 🚀\n\n**What is it about:**\n\n📌 In this [notebook]( https://www.kaggle.com/code/ivanisaev/model-based-on-previous-predictions) I wrote about how to train model on correct labels for other questions in current session. \n\n📌 For this notebook I similarly trained a meta-XGBoost model but instead of correct labels for training I used predictions (probabilities not labels after converting by threshold) of this model on trained dataset. And I use this meta-model to mix it with main model to obtain final predictions.\n\n**My approach**\n\n📌 As a main model I used model from [this notebook](https://www.kaggle.com/code/vadimkamaev/catboost-mix) and trained the meta-model using predictions of this main model on train data.\n\n📌☝️ In the iter test I save predictions for the previous questions for current and previous level groups. This previous predictions are used in final mix by meta-model to make final predictions more robust.\n \n📌 To mix predictions of main Catboost-mix model and meta-model I used similar schema as in [my previous notebook](https://www.kaggle.com/code/ivanisaev/0-697-ensamble-nn-catboost). \n\n📌 Again I didn't receive score improvement in public LB but performance didn't decrease. I hope this is a good sign that model become tore robust for private LB data.\n\nWill appreciate your feedback! Thank you and have a nice day!🙂",
    "2292696": "Thanks for the notebook."
  },
  "source": "meta"
}