{
  "id": 55766,
  "title": "New Features -> New Hyperparameters",
  "url": "/competitions/talkingdata-adtracking-fraud-detection/discussion/55766",
  "author_name": "",
  "post_date": "2018-05-01T12:40:52.943422500Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi,\nDo you tune your hyperparameters for new features?\nLet's say I have a model with specific hyperparameters, and my max_depth is low \ne.g 4 \nAdding 5 new features to the model might cause an unexpected outcome.\nLet's say that my 5 new features are good for the model, but are not used properly by the model because the depth is low, and for them specifically, maybe they need more depth.</p>\n\n<p>In this case i can see no improvement in the model at all although if i would have changed the hyperparameters i would discover they do benefit.</p>\n\n<p>This could be same for number of leaves or col.sample also..</p>\n\n<p>How do you handle adding more features in regards to that?</p>",
  "messages": [
    {
      "id": "321500",
      "postDate": "05/01/2018 12:40:52",
      "content": "<p>Hi,\nDo you tune your hyperparameters for new features?\nLet's say I have a model with specific hyperparameters, and my max_depth is low \ne.g 4 \nAdding 5 new features to the model might cause an unexpected outcome.\nLet's say that my 5 new features are good for the model, but are not used properly by the model because the depth is low, and for them specifically, maybe they need more depth.</p>\n\n<p>In this case i can see no improvement in the model at all although if i would have changed the hyperparameters i would discover they do benefit.</p>\n\n<p>This could be same for number of leaves or col.sample also..</p>\n\n<p>How do you handle adding more features in regards to that?</p>",
      "rawMarkdown": "Hi,\nDo you tune your hyperparameters for new features?\nLet's say I have a model with specific hyperparameters, and my max_depth is low \ne.g 4 \nAdding 5 new features to the model might cause an unexpected outcome.\nLet's say that my 5 new features are good for the model, but are not used properly by the model because the depth is low, and for them specifically, maybe they need more depth.\n\nIn this case i can see no improvement in the model at all although if i would have changed the hyperparameters i would discover they do benefit.\n\nThis could be same for number of leaves or col.sample also..\n\nHow do you handle adding more features in regards to that?",
      "votes": null
    },
    {
      "id": "321632",
      "postDate": "05/01/2018 17:25:31",
      "content": "<blockquote>\n  <p>Do you tune your hyperparameters for new features?</p>\n</blockquote>\n\n<p>In general, yes. Given the size of this dataset, it may not be practical to tune them from the scratch. I'd make small changes up and down from your existing parameters, giving priority to <code>['subsample' , 'colsample_bytree', 'min_child_weight', 'scale_pos_weight']</code>. Maybe <code>num_leaves</code> as well. I would not expect <code>max_depth</code> to change.</p>",
      "rawMarkdown": "&gt; Do you tune your hyperparameters for new features?\n\nIn general, yes. Given the size of this dataset, it may not be practical to tune them from the scratch. I'd make small changes up and down from your existing parameters, giving priority to `['subsample' , 'colsample_bytree', 'min_child_weight', 'scale_pos_weight']`. Maybe `num_leaves` as well. I would not expect `max_depth` to change.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 321632,
      "author_name": "tilii7",
      "author_url": "",
      "post_date": "05/01/2018 17:25:31",
      "content": "<blockquote>\n  <p>Do you tune your hyperparameters for new features?</p>\n</blockquote>\n\n<p>In general, yes. Given the size of this dataset, it may not be practical to tune them from the scratch. I'd make small changes up and down from your existing parameters, giving priority to <code>['subsample' , 'colsample_bytree', 'min_child_weight', 'scale_pos_weight']</code>. Maybe <code>num_leaves</code> as well. I would not expect <code>max_depth</code> to change.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "321500": "Hi,\nDo you tune your hyperparameters for new features?\nLet's say I have a model with specific hyperparameters, and my max_depth is low \ne.g 4 \nAdding 5 new features to the model might cause an unexpected outcome.\nLet's say that my 5 new features are good for the model, but are not used properly by the model because the depth is low, and for them specifically, maybe they need more depth.\n\nIn this case i can see no improvement in the model at all although if i would have changed the hyperparameters i would discover they do benefit.\n\nThis could be same for number of leaves or col.sample also..\n\nHow do you handle adding more features in regards to that?",
    "321632": "&gt; Do you tune your hyperparameters for new features?\n\nIn general, yes. Given the size of this dataset, it may not be practical to tune them from the scratch. I'd make small changes up and down from your existing parameters, giving priority to `['subsample' , 'colsample_bytree', 'min_child_weight', 'scale_pos_weight']`. Maybe `num_leaves` as well. I would not expect `max_depth` to change."
  },
  "source": "meta"
}