{
  "id": 74746,
  "title": "Smote didn't help with neural networks",
  "url": "/competitions/PLAsTiCC-2018/discussion/74746",
  "author_name": "",
  "post_date": "2018-12-15T01:58:03.597411100Z",
  "votes": 3,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Interestingly enough after struggling to integrate Smote with neural networks (had to convert the y back and forth to/from categorical) I found that it actually made my NN score worse.  Glad it helped for LGBM.  Any ideas why this might be the case?</p>",
  "messages": [
    {
      "id": "439244",
      "postDate": "12/15/2018 01:58:03",
      "content": "<p>Interestingly enough after struggling to integrate Smote with neural networks (had to convert the y back and forth to/from categorical) I found that it actually made my NN score worse.  Glad it helped for LGBM.  Any ideas why this might be the case?</p>",
      "rawMarkdown": "Interestingly enough after struggling to integrate Smote with neural networks (had to convert the y back and forth to/from categorical) I found that it actually made my NN score worse.  Glad it helped for LGBM.  Any ideas why this might be the case?",
      "votes": null
    },
    {
      "id": "440051",
      "postDate": "12/17/2018 00:47:25",
      "content": "<p>I've only worked on my NN for a few hours, but so far SMOTE has been a savior because my model was outputting garbage due to the imbalance between classes...</p>",
      "rawMarkdown": "I've only worked on my NN for a few hours, but so far SMOTE has been a savior because my model was outputting garbage due to the imbalance between classes...",
      "votes": null
    },
    {
      "id": "440321",
      "postDate": "12/17/2018 11:59:00",
      "content": "<p>yes that seems to be true.I am trying very hard to train a nn model on the same features that I used in lgb. results are as follows:\nlgb- CV scored 0.577x and lb score was 0.963\nnn- CV scored 0.71x and I didn't checked lb because of poor performance.</p>\n\n<p>Any hints, what i am missing??</p>",
      "rawMarkdown": "yes that seems to be true.I am trying very hard to train a nn model on the same features that I used in lgb. results are as follows:\nlgb- CV scored 0.577x and lb score was 0.963\nnn- CV scored 0.71x and I didn't checked lb because of poor performance.\n\nAny hints, what i am missing??",
      "votes": null
    },
    {
      "id": "440485",
      "postDate": "12/17/2018 15:57:51",
      "content": "<p>Probably setting a lower rate in SMOTE for imbalanced cases will help. Take into account that SMOTE is distorting variable distribution as it generates new synthetic cases. A good idea would be, after SMOTE, take a variable explanation test (e.g., mutual information or the like) in order to see how much variable distributions shifted.</p>",
      "rawMarkdown": "Probably setting a lower rate in SMOTE for imbalanced cases will help. Take into account that SMOTE is distorting variable distribution as it generates new synthetic cases. A good idea would be, after SMOTE, take a variable explanation test (e.g., mutual information or the like) in order to see how much variable distributions shifted.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 440051,
      "author_name": "sdoria",
      "author_url": "",
      "post_date": "12/17/2018 00:47:25",
      "content": "<p>I've only worked on my NN for a few hours, but so far SMOTE has been a savior because my model was outputting garbage due to the imbalance between classes...</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 440321,
      "author_name": "aashish7936",
      "author_url": "",
      "post_date": "12/17/2018 11:59:00",
      "content": "<p>yes that seems to be true.I am trying very hard to train a nn model on the same features that I used in lgb. results are as follows:\nlgb- CV scored 0.577x and lb score was 0.963\nnn- CV scored 0.71x and I didn't checked lb because of poor performance.</p>\n\n<p>Any hints, what i am missing??</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 440485,
      "author_name": "andreusancho",
      "author_url": "",
      "post_date": "12/17/2018 15:57:51",
      "content": "<p>Probably setting a lower rate in SMOTE for imbalanced cases will help. Take into account that SMOTE is distorting variable distribution as it generates new synthetic cases. A good idea would be, after SMOTE, take a variable explanation test (e.g., mutual information or the like) in order to see how much variable distributions shifted.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "439244": "Interestingly enough after struggling to integrate Smote with neural networks (had to convert the y back and forth to/from categorical) I found that it actually made my NN score worse.  Glad it helped for LGBM.  Any ideas why this might be the case?",
    "440051": "I've only worked on my NN for a few hours, but so far SMOTE has been a savior because my model was outputting garbage due to the imbalance between classes...",
    "440321": "yes that seems to be true.I am trying very hard to train a nn model on the same features that I used in lgb. results are as follows:\nlgb- CV scored 0.577x and lb score was 0.963\nnn- CV scored 0.71x and I didn't checked lb because of poor performance.\n\nAny hints, what i am missing??",
    "440485": "Probably setting a lower rate in SMOTE for imbalanced cases will help. Take into account that SMOTE is distorting variable distribution as it generates new synthetic cases. A good idea would be, after SMOTE, take a variable explanation test (e.g., mutual information or the like) in order to see how much variable distributions shifted."
  },
  "source": "meta"
}