{
  "id": 55919,
  "title": "LightGBM stops too early",
  "url": "/competitions/talkingdata-adtracking-fraud-detection/discussion/55919",
  "author_name": "",
  "post_date": "2018-05-03T05:36:48.924231Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I ran my model using LightGBM, I set the round = 1000,  but it stop at round 70 and show that the best iteration is 41. I worried that my model didn't get fully - trained. This is weird, I want to get more advice. Can you help me?</p>",
  "messages": [
    {
      "id": "322512",
      "postDate": "05/03/2018 05:36:48",
      "content": "<p>I ran my model using LightGBM, I set the round = 1000,  but it stop at round 70 and show that the best iteration is 41. I worried that my model didn't get fully - trained. This is weird, I want to get more advice. Can you help me?</p>",
      "rawMarkdown": "I ran my model using LightGBM, I set the round = 1000,  but it stop at round 70 and show that the best iteration is 41. I worried that my model didn't get fully - trained. This is weird, I want to get more advice. Can you help me?",
      "votes": null
    },
    {
      "id": "322516",
      "postDate": "05/03/2018 05:49:50",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/nogiveup\">poison code</a>, did you put set early_stopping_rounds\nTypically set it to 30 or 50\nAlso, what is your learning rate\nyou could set it to 0.1</p>",
      "rawMarkdown": "Hi [poison code](https://www.kaggle.com/nogiveup), did you put set early_stopping_rounds\nTypically set it to 30 or 50\nAlso, what is your learning rate\nyou could set it to 0.1",
      "votes": null
    },
    {
      "id": "322518",
      "postDate": "05/03/2018 05:50:14",
      "content": "<p>Hello, increase your early stopping parametr.</p>",
      "rawMarkdown": "Hello, increase your early stopping parametr.",
      "votes": null
    },
    {
      "id": "322757",
      "postDate": "05/03/2018 15:49:31",
      "content": "<p>You are probably using the parameter \"early_stopping_rounds\".\nIf set to e.g 30, it will keep training until for 30 consecutive rounds the performance of the validation does not get better, and then stop.\nTo maybe increase your score up a bit, use a low learning_rate, e.g 0.1, with a bit higher \"early_stopping_rounds\", e.g 50\n(it will take longer to run though)</p>\n\n<p>Also - you can set which data it will use to validate \"early_stopping_rounds\" with the parameter \"valids\", it's better you use a cross validation set there rather than your training set.</p>",
      "rawMarkdown": "You are probably using the parameter \"early_stopping_rounds\".\nIf set to e.g 30, it will keep training until for 30 consecutive rounds the performance of the validation does not get better, and then stop.\nTo maybe increase your score up a bit, use a low learning_rate, e.g 0.1, with a bit higher \"early_stopping_rounds\", e.g 50\n(it will take longer to run though)\n\nAlso - you can set which data it will use to validate \"early_stopping_rounds\" with the parameter \"valids\", it's better you use a cross validation set there rather than your training set.",
      "votes": null
    },
    {
      "id": "322799",
      "postDate": "05/03/2018 17:15:44",
      "content": "<p>Thank you very much.  </p>",
      "rawMarkdown": "Thank you very much.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 322516,
      "author_name": "ericbenhamou",
      "author_url": "",
      "post_date": "05/03/2018 05:49:50",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/nogiveup\">poison code</a>, did you put set early_stopping_rounds\nTypically set it to 30 or 50\nAlso, what is your learning rate\nyou could set it to 0.1</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 322518,
      "author_name": "izmaylov",
      "author_url": "",
      "post_date": "05/03/2018 05:50:14",
      "content": "<p>Hello, increase your early stopping parametr.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 322757,
      "author_name": "tpthegreat",
      "author_url": "",
      "post_date": "05/03/2018 15:49:31",
      "content": "<p>You are probably using the parameter \"early_stopping_rounds\".\nIf set to e.g 30, it will keep training until for 30 consecutive rounds the performance of the validation does not get better, and then stop.\nTo maybe increase your score up a bit, use a low learning_rate, e.g 0.1, with a bit higher \"early_stopping_rounds\", e.g 50\n(it will take longer to run though)</p>\n\n<p>Also - you can set which data it will use to validate \"early_stopping_rounds\" with the parameter \"valids\", it's better you use a cross validation set there rather than your training set.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 322799,
      "author_name": "nogiveup",
      "author_url": "",
      "post_date": "05/03/2018 17:15:44",
      "content": "<p>Thank you very much.  </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "322512": "I ran my model using LightGBM, I set the round = 1000,  but it stop at round 70 and show that the best iteration is 41. I worried that my model didn't get fully - trained. This is weird, I want to get more advice. Can you help me?",
    "322516": "Hi [poison code](https://www.kaggle.com/nogiveup), did you put set early_stopping_rounds\nTypically set it to 30 or 50\nAlso, what is your learning rate\nyou could set it to 0.1",
    "322518": "Hello, increase your early stopping parametr.",
    "322757": "You are probably using the parameter \"early_stopping_rounds\".\nIf set to e.g 30, it will keep training until for 30 consecutive rounds the performance of the validation does not get better, and then stop.\nTo maybe increase your score up a bit, use a low learning_rate, e.g 0.1, with a bit higher \"early_stopping_rounds\", e.g 50\n(it will take longer to run though)\n\nAlso - you can set which data it will use to validate \"early_stopping_rounds\" with the parameter \"valids\", it's better you use a cross validation set there rather than your training set.",
    "322799": "Thank you very much."
  },
  "source": "meta"
}