{
  "id": 14661,
  "title": "How do you deal with class imbalance?",
  "url": "/competitions/facebook-recruiting-iv-human-or-bot/discussion/14661",
  "author_name": "",
  "post_date": "2015-06-10T11:58:49.143Z",
  "votes": null,
  "comment_count": 7,
  "views": 1650,
  "content": "<p>Hi all,</p>\n\n<p>Congratulations on the winners!</p>\n<p>For most of time I was stuck at model's under performance, where there are many false-positives.</p>\n<p>I was attributing this to the class imbalance, and applied SMOTE and stratified sampling to 5~10 folds.</p>\n<p>It helped a bit, but not much.</p>\n\n<p>As discussed in another topic, feature engineering matters the most.</p>\n<p>However, how do you deal with the class imbalance?</p>\n<p>Is it a trivial thing in this competition?</p>\n\n<p>Thanks for all the sharing,</p>\n<p>Sunny</p>",
  "messages": [
    {
      "id": "81473",
      "postDate": "06/10/2015 11:58:49",
      "content": "<p>Hi all,</p>\n\n<p>Congratulations on the winners!</p>\n<p>For most of time I was stuck at model's under performance, where there are many false-positives.</p>\n<p>I was attributing this to the class imbalance, and applied SMOTE and stratified sampling to 5~10 folds.</p>\n<p>It helped a bit, but not much.</p>\n\n<p>As discussed in another topic, feature engineering matters the most.</p>\n<p>However, how do you deal with the class imbalance?</p>\n<p>Is it a trivial thing in this competition?</p>\n\n<p>Thanks for all the sharing,</p>\n<p>Sunny</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81479",
      "postDate": "06/10/2015 13:14:28",
      "content": "<p>In my submissions&nbsp;Borderline-SMOTE seemed to help. SMOTE was helpful too but did not produce highly significant improvements.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81560",
      "postDate": "06/11/2015 02:44:28",
      "content": "<p>Hi Sushant,</p>\n\n<p>How much did Borderline-SMOTE helped on AUC?</p>\n\n<p>[quote=Sushant Tripathy;81479]</p>\n<p>In my submissions&nbsp;Borderline-SMOTE seemed to help. SMOTE was helpful too but did not produce highly significant improvements.</p>\n<p>[/quote]</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81562",
      "postDate": "06/11/2015 02:50:45",
      "content": "<p>My CV score went up by 0.005, that difference was also reflected on the private and public LB scores. Not a lot, but a little goes a long way here :)</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81564",
      "postDate": "06/11/2015 02:51:00",
      "content": "<p>Oops double post!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81568",
      "postDate": "06/11/2015 04:01:58",
      "content": "<p>[quote=Sushant Tripathy;81562]</p>\n<p>My CV score went up by 0.005, that difference was also reflected on the private and public LB scores. Not a lot, but a little goes a long way here :)</p>\n<p>[/quote]</p>\n\n<p>Good to know, thanks for sharing!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81570",
      "postDate": "06/11/2015 04:16:44",
      "content": "<p>BTW, did you applied Borderline-SMOTE1 or Borderline-SMOTE2?</p>\n\n<p>[quote=Sushant Tripathy;81562]</p>\n<p>My CV score went up by 0.005, that difference was also reflected on the private and public LB scores. Not a lot, but a little goes a long way here :)</p>\n<p>[/quote]</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81572",
      "postDate": "06/11/2015 04:36:02",
      "content": "<p>[quote=bluesilence;81570]</p>\n<p>BTW, did you applied Borderline-SMOTE1 or Borderline-SMOTE2?</p>\n<p>[quote=Sushant Tripathy;81562]</p>\n<p>My CV score went up by 0.005, that difference was also reflected on the private and public LB scores. Not a lot, but a little goes a long way here :)</p>\n<p>[/quote]</p>\n<p>[/quote]</p>\n<p>:)&nbsp;Glad you found the information useful;&nbsp;I used Borderline-SMOTE1 (that produced the best results during CV). A point to note was that with bSMOTE1 I could push the AUC score of a 2 hidden layer NN (60,60 hidden layers) to 0.90xx...close to 0.91.</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 81479,
      "author_name": "sushanttripathy",
      "author_url": "",
      "post_date": "06/10/2015 13:14:28",
      "content": "<p>In my submissions&nbsp;Borderline-SMOTE seemed to help. SMOTE was helpful too but did not produce highly significant improvements.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81560,
      "author_name": "bluesilence",
      "author_url": "",
      "post_date": "06/11/2015 02:44:28",
      "content": "<p>Hi Sushant,</p>\n\n<p>How much did Borderline-SMOTE helped on AUC?</p>\n\n<p>[quote=Sushant Tripathy;81479]</p>\n<p>In my submissions&nbsp;Borderline-SMOTE seemed to help. SMOTE was helpful too but did not produce highly significant improvements.</p>\n<p>[/quote]</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81562,
      "author_name": "sushanttripathy",
      "author_url": "",
      "post_date": "06/11/2015 02:50:45",
      "content": "<p>My CV score went up by 0.005, that difference was also reflected on the private and public LB scores. Not a lot, but a little goes a long way here :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81564,
      "author_name": "sushanttripathy",
      "author_url": "",
      "post_date": "06/11/2015 02:51:00",
      "content": "<p>Oops double post!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81568,
      "author_name": "bluesilence",
      "author_url": "",
      "post_date": "06/11/2015 04:01:58",
      "content": "<p>[quote=Sushant Tripathy;81562]</p>\n<p>My CV score went up by 0.005, that difference was also reflected on the private and public LB scores. Not a lot, but a little goes a long way here :)</p>\n<p>[/quote]</p>\n\n<p>Good to know, thanks for sharing!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81570,
      "author_name": "bluesilence",
      "author_url": "",
      "post_date": "06/11/2015 04:16:44",
      "content": "<p>BTW, did you applied Borderline-SMOTE1 or Borderline-SMOTE2?</p>\n\n<p>[quote=Sushant Tripathy;81562]</p>\n<p>My CV score went up by 0.005, that difference was also reflected on the private and public LB scores. Not a lot, but a little goes a long way here :)</p>\n<p>[/quote]</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81572,
      "author_name": "sushanttripathy",
      "author_url": "",
      "post_date": "06/11/2015 04:36:02",
      "content": "<p>[quote=bluesilence;81570]</p>\n<p>BTW, did you applied Borderline-SMOTE1 or Borderline-SMOTE2?</p>\n<p>[quote=Sushant Tripathy;81562]</p>\n<p>My CV score went up by 0.005, that difference was also reflected on the private and public LB scores. Not a lot, but a little goes a long way here :)</p>\n<p>[/quote]</p>\n<p>[/quote]</p>\n<p>:)&nbsp;Glad you found the information useful;&nbsp;I used Borderline-SMOTE1 (that produced the best results during CV). A point to note was that with bSMOTE1 I could push the AUC score of a 2 hidden layer NN (60,60 hidden layers) to 0.90xx...close to 0.91.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "81473": "",
    "81479": "",
    "81560": "",
    "81562": "",
    "81564": "",
    "81568": "",
    "81570": "",
    "81572": ""
  },
  "source": "meta"
}