{
  "id": 174743,
  "title": "Binary Cross Entropy and Focal Loss Ensemble",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/174743",
  "author_name": "",
  "post_date": "2020-08-15T05:42:23.515861100Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi all! I was wondering if anyone has gotten decent results ensembling models using different loss functions (specifically binary and focal loss)? For me, it seems that there hasn't really been much improvement and ensembling using the same loss function might be the way to go.</p>",
  "messages": [
    {
      "id": "971028",
      "postDate": "08/15/2020 05:42:23",
      "content": "<p>Hi all! I was wondering if anyone has gotten decent results ensembling models using different loss functions (specifically binary and focal loss)? For me, it seems that there hasn't really been much improvement and ensembling using the same loss function might be the way to go.</p>",
      "rawMarkdown": "Hi all! I was wondering if anyone has gotten decent results ensembling models using different loss functions (specifically binary and focal loss)? For me, it seems that there hasn't really been much improvement and ensembling using the same loss function might be the way to go.",
      "votes": null
    },
    {
      "id": "971547",
      "postDate": "08/15/2020 16:56:30",
      "content": "<p>I've seen a lot of folks comment that they have had success with BCE but not with Focal Loss. </p>\n<p>Does anyone know why Focal Loss might not be effective here?</p>",
      "rawMarkdown": "I've seen a lot of folks comment that they have had success with BCE but not with Focal Loss. \n\nDoes anyone know why Focal Loss might not be effective here?",
      "votes": null
    },
    {
      "id": "971948",
      "postDate": "08/16/2020 05:39:38",
      "content": "<p>I don't know why it wouldn't be effective. I would say that it should be as effective if not more than BCE. Atleast, I have gotten better CV scores with focal loss. I have a feeling that BCE might be more sensitive to small adjustments perhaps giving it better CV when ensembling since the range of values predicted by BCE is largely skewed heavily towards the negative class.</p>",
      "rawMarkdown": "I don't know why it wouldn't be effective. I would say that it should be as effective if not more than BCE. Atleast, I have gotten better CV scores with focal loss. I have a feeling that BCE might be more sensitive to small adjustments perhaps giving it better CV when ensembling since the range of values predicted by BCE is largely skewed heavily towards the negative class.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 971547,
      "author_name": "campbellhutcheson",
      "author_url": "",
      "post_date": "08/15/2020 16:56:30",
      "content": "<p>I've seen a lot of folks comment that they have had success with BCE but not with Focal Loss. </p>\n<p>Does anyone know why Focal Loss might not be effective here?</p>",
      "votes": null,
      "replies": [
        {
          "id": 971948,
          "author_name": "richardxiao03",
          "author_url": "",
          "post_date": "08/16/2020 05:39:38",
          "content": "<p>I don't know why it wouldn't be effective. I would say that it should be as effective if not more than BCE. Atleast, I have gotten better CV scores with focal loss. I have a feeling that BCE might be more sensitive to small adjustments perhaps giving it better CV when ensembling since the range of values predicted by BCE is largely skewed heavily towards the negative class.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "971028": "Hi all! I was wondering if anyone has gotten decent results ensembling models using different loss functions (specifically binary and focal loss)? For me, it seems that there hasn't really been much improvement and ensembling using the same loss function might be the way to go.",
    "971547": "I've seen a lot of folks comment that they have had success with BCE but not with Focal Loss. \n\nDoes anyone know why Focal Loss might not be effective here?",
    "971948": "I don't know why it wouldn't be effective. I would say that it should be as effective if not more than BCE. Atleast, I have gotten better CV scores with focal loss. I have a feeling that BCE might be more sensitive to small adjustments perhaps giving it better CV when ensembling since the range of values predicted by BCE is largely skewed heavily towards the negative class."
  },
  "source": "meta"
}