{
  "id": 452274,
  "title": "Loss functions discussion",
  "url": "/competitions/open-problems-single-cell-perturbations/discussion/452274",
  "author_name": "",
  "post_date": "2023-11-01T14:48:22.398066100Z",
  "votes": 6,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Hello everyone,</p>\n<p>I'm currently seeking insights into which loss function might yield better performance. In my experiments, I utilized MAE, which delivered remarkable results compared to a custom MRRMSE. I'm considering experimenting with the Huber loss function, as I'm incorporating target encoding features for each cell_type and sm_name. Has anyone faced any challenges related to outliers while encoding these additional features?</p>",
  "messages": [
    {
      "id": "2508203",
      "postDate": "11/01/2023 14:48:22",
      "content": "<p>Hello everyone,</p>\n<p>I'm currently seeking insights into which loss function might yield better performance. In my experiments, I utilized MAE, which delivered remarkable results compared to a custom MRRMSE. I'm considering experimenting with the Huber loss function, as I'm incorporating target encoding features for each cell_type and sm_name. Has anyone faced any challenges related to outliers while encoding these additional features?</p>",
      "rawMarkdown": "Hello everyone,\n\nI'm currently seeking insights into which loss function might yield better performance. In my experiments, I utilized MAE, which delivered remarkable results compared to a custom MRRMSE. I'm considering experimenting with the Huber loss function, as I'm incorporating target encoding features for each cell_type and sm_name. Has anyone faced any challenges related to outliers while encoding these additional features?",
      "votes": null
    },
    {
      "id": "2508499",
      "postDate": "11/01/2023 18:54:37",
      "content": "<p>MAE has performed better for me than MSE. I've been working with mainly Pytorch NNs. I played around with a MRRMSE loss function at one point but didn't test it extensively.</p>",
      "rawMarkdown": "MAE has performed better for me than MSE. I've been working with mainly Pytorch NNs. I played around with a MRRMSE loss function at one point but didn't test it extensively.",
      "votes": null
    },
    {
      "id": "2509127",
      "postDate": "11/02/2023 07:27:47",
      "content": "<p>Are you regressing on the raw target value? Or you have done some standard scaling and dimensionality reductions?</p>",
      "rawMarkdown": "Are you regressing on the raw target value? Or you have done some standard scaling and dimensionality reductions?",
      "votes": null
    },
    {
      "id": "2509136",
      "postDate": "11/02/2023 07:35:02",
      "content": "<p>I experimented with all three options, employing cross-validation to determine the optimal dimension, or alternatively, sticking with the unprocessed target values.</p>",
      "rawMarkdown": "I experimented with all three options, employing cross-validation to determine the optimal dimension, or alternatively, sticking with the unprocessed target values.",
      "votes": null
    },
    {
      "id": "2509152",
      "postDate": "11/02/2023 07:52:40",
      "content": "<p>Thanks for your reply! From my side, I found that dimension reduction is not very helpful when using mrrmse. I guess that's why mae works well.</p>",
      "rawMarkdown": "Thanks for your reply! From my side, I found that dimension reduction is not very helpful when using mrrmse. I guess that's why mae works well.",
      "votes": null
    },
    {
      "id": "2509384",
      "postDate": "11/02/2023 10:55:10",
      "content": "<p>I recommend giving the Huber loss a try and would be interested to hear if it has a positive impact on your model</p>",
      "rawMarkdown": "I recommend giving the Huber loss a try and would be interested to hear if it has a positive impact on your model",
      "votes": null
    },
    {
      "id": "2510602",
      "postDate": "11/03/2023 06:03:46",
      "content": "<p>Yes! I will experiment with this more.</p>",
      "rawMarkdown": "Yes! I will experiment with this more.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2508499,
      "author_name": "mikelong",
      "author_url": "",
      "post_date": "11/01/2023 18:54:37",
      "content": "<p>MAE has performed better for me than MSE. I've been working with mainly Pytorch NNs. I played around with a MRRMSE loss function at one point but didn't test it extensively.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2509127,
      "author_name": "superdanielshao",
      "author_url": "",
      "post_date": "11/02/2023 07:27:47",
      "content": "<p>Are you regressing on the raw target value? Or you have done some standard scaling and dimensionality reductions?</p>",
      "votes": null,
      "replies": [
        {
          "id": 2509136,
          "author_name": "eliork",
          "author_url": "",
          "post_date": "11/02/2023 07:35:02",
          "content": "<p>I experimented with all three options, employing cross-validation to determine the optimal dimension, or alternatively, sticking with the unprocessed target values.</p>",
          "votes": null,
          "replies": [
            {
              "id": 2509152,
              "author_name": "superdanielshao",
              "author_url": "",
              "post_date": "11/02/2023 07:52:40",
              "content": "<p>Thanks for your reply! From my side, I found that dimension reduction is not very helpful when using mrrmse. I guess that's why mae works well.</p>",
              "votes": null,
              "replies": [
                {
                  "id": 2509384,
                  "author_name": "eliork",
                  "author_url": "",
                  "post_date": "11/02/2023 10:55:10",
                  "content": "<p>I recommend giving the Huber loss a try and would be interested to hear if it has a positive impact on your model</p>",
                  "votes": null,
                  "replies": [
                    {
                      "id": 2510602,
                      "author_name": "superdanielshao",
                      "author_url": "",
                      "post_date": "11/03/2023 06:03:46",
                      "content": "<p>Yes! I will experiment with this more.</p>",
                      "votes": null,
                      "replies": []
                    }
                  ]
                }
              ]
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2508203": "Hello everyone,\n\nI'm currently seeking insights into which loss function might yield better performance. In my experiments, I utilized MAE, which delivered remarkable results compared to a custom MRRMSE. I'm considering experimenting with the Huber loss function, as I'm incorporating target encoding features for each cell_type and sm_name. Has anyone faced any challenges related to outliers while encoding these additional features?",
    "2508499": "MAE has performed better for me than MSE. I've been working with mainly Pytorch NNs. I played around with a MRRMSE loss function at one point but didn't test it extensively.",
    "2509127": "Are you regressing on the raw target value? Or you have done some standard scaling and dimensionality reductions?",
    "2509136": "I experimented with all three options, employing cross-validation to determine the optimal dimension, or alternatively, sticking with the unprocessed target values.",
    "2509152": "Thanks for your reply! From my side, I found that dimension reduction is not very helpful when using mrrmse. I guess that's why mae works well.",
    "2509384": "I recommend giving the Huber loss a try and would be interested to hear if it has a positive impact on your model",
    "2510602": "Yes! I will experiment with this more."
  },
  "source": "meta"
}