{
  "id": 21433,
  "title": "Question regarding xgboost's ETA and num_rounds",
  "url": "/competitions/facebook-v-predicting-check-ins/discussion/21433",
  "author_name": "",
  "post_date": "2016-06-04T19:52:59.687Z",
  "votes": null,
  "comment_count": 3,
  "views": 815,
  "content": "<p>Hi all:</p>\n\n<p>I'm new to xgboost and trying to learn. Seems like ETA is a multiplier applied to outputs of each boosters, and num_rounds is the number of boosters we want for the entire model. Hence my feeling is that, for a large ETA value, we want to use a  smaller num_rounds value, and vice versa.</p>\n\n<p>Question is, what would be the optimal ratio of these two? Say I'm using ETA = 0.1, hence I'm taking one tenth the magnitude of the value generated from each booster; then for num_rounds, should I use a value that is greater or equal to ten?</p>\n\n<p>Or are there more factors that I should take into consideration as well? Any help is much appreciated!</p>",
  "messages": [
    {
      "id": "122522",
      "postDate": "06/04/2016 19:52:59",
      "content": "<p>Hi all:</p>\n\n<p>I'm new to xgboost and trying to learn. Seems like ETA is a multiplier applied to outputs of each boosters, and num_rounds is the number of boosters we want for the entire model. Hence my feeling is that, for a large ETA value, we want to use a  smaller num_rounds value, and vice versa.</p>\n\n<p>Question is, what would be the optimal ratio of these two? Say I'm using ETA = 0.1, hence I'm taking one tenth the magnitude of the value generated from each booster; then for num_rounds, should I use a value that is greater or equal to ten?</p>\n\n<p>Or are there more factors that I should take into consideration as well? Any help is much appreciated!</p>",
      "rawMarkdown": "Hi all:\r\n\r\nI'm new to xgboost and trying to learn. Seems like ETA is a multiplier applied to outputs of each boosters, and num_rounds is the number of boosters we want for the entire model. Hence my feeling is that, for a large ETA value, we want to use a  smaller num_rounds value, and vice versa.\r\n\r\nQuestion is, what would be the optimal ratio of these two? Say I'm using ETA = 0.1, hence I'm taking one tenth the magnitude of the value generated from each booster; then for num_rounds, should I use a value that is greater or equal to ten?\r\n\r\nOr are there more factors that I should take into consideration as well? Any help is much appreciated!",
      "votes": null
    },
    {
      "id": "122525",
      "postDate": "06/04/2016 20:13:57",
      "content": "<p>Hi Jasoos, check out this link for hyperparameter tuning in XGBoost models: <a href=\"http://datascience.stackexchange.com/questions/9364/hypertuning-xgboost-parameters\">http://datascience.stackexchange.com/questions/9364/hypertuning-xgboost-parameters</a></p>",
      "rawMarkdown": "Hi Jasoos, check out this link for hyperparameter tuning in XGBoost models: http://datascience.stackexchange.com/questions/9364/hypertuning-xgboost-parameters",
      "votes": null
    },
    {
      "id": "122526",
      "postDate": "06/04/2016 20:26:25",
      "content": "<p>Thanks a lot @Tom, and congratulations!</p>",
      "rawMarkdown": "Thanks a lot @Tom, and congratulations!",
      "votes": null
    },
    {
      "id": "122528",
      "postDate": "06/04/2016 20:50:17",
      "content": "<p>You're welcome and thank you but it's still very early in the competition so I guess we'll be seeing quite some changes at the top of the leaderboard!</p>",
      "rawMarkdown": "You're welcome and thank you but it's still very early in the competition so I guess we'll be seeing quite some changes at the top of the leaderboard!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 122525,
      "author_name": "tvdwiele",
      "author_url": "",
      "post_date": "06/04/2016 20:13:57",
      "content": "<p>Hi Jasoos, check out this link for hyperparameter tuning in XGBoost models: <a href=\"http://datascience.stackexchange.com/questions/9364/hypertuning-xgboost-parameters\">http://datascience.stackexchange.com/questions/9364/hypertuning-xgboost-parameters</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 122526,
      "author_name": "howardxie",
      "author_url": "",
      "post_date": "06/04/2016 20:26:25",
      "content": "<p>Thanks a lot @Tom, and congratulations!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 122528,
      "author_name": "tvdwiele",
      "author_url": "",
      "post_date": "06/04/2016 20:50:17",
      "content": "<p>You're welcome and thank you but it's still very early in the competition so I guess we'll be seeing quite some changes at the top of the leaderboard!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "122522": "Hi all:\r\n\r\nI'm new to xgboost and trying to learn. Seems like ETA is a multiplier applied to outputs of each boosters, and num_rounds is the number of boosters we want for the entire model. Hence my feeling is that, for a large ETA value, we want to use a  smaller num_rounds value, and vice versa.\r\n\r\nQuestion is, what would be the optimal ratio of these two? Say I'm using ETA = 0.1, hence I'm taking one tenth the magnitude of the value generated from each booster; then for num_rounds, should I use a value that is greater or equal to ten?\r\n\r\nOr are there more factors that I should take into consideration as well? Any help is much appreciated!",
    "122525": "Hi Jasoos, check out this link for hyperparameter tuning in XGBoost models: http://datascience.stackexchange.com/questions/9364/hypertuning-xgboost-parameters",
    "122526": "Thanks a lot @Tom, and congratulations!",
    "122528": "You're welcome and thank you but it's still very early in the competition so I guess we'll be seeing quite some changes at the top of the leaderboard!"
  },
  "source": "meta"
}