{
  "id": 406139,
  "title": "How should we address the imbalance dataset issue in this competition? And should we model this a classification problem or a regression problem?",
  "url": "/competitions/tlvmc-parkinsons-freezing-gait-prediction/discussion/406139",
  "author_name": "",
  "post_date": "2023-05-01T03:21:53.480221800Z",
  "votes": 7,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Hi,</p>\n<p>I have some questions about this competition, and I hope there is anyone who could help with this :)</p>\n<p>1) How should we address the imbalance dataset issue in this competition? </p>\n<p>I tried a couple of approaches like data up-sampling and weighted loss. Still, I got a very low precision score on the activity Start Hesitation and Walking :(  For example, I got an average score 0.272, and 0.077,0.687,0.052 for  Start Hesitation, Turn, and Walking.</p>\n<p>2) is this a classification problem or a regression problem?</p>\n<p>I see many notebooks model this as a regression problem. Is there an issue if we model it as a classification problem?</p>\n<p>Thank you</p>",
  "messages": [
    {
      "id": "2240950",
      "postDate": "05/01/2023 03:21:53",
      "content": "<p>Hi,</p>\n<p>I have some questions about this competition, and I hope there is anyone who could help with this :)</p>\n<p>1) How should we address the imbalance dataset issue in this competition? </p>\n<p>I tried a couple of approaches like data up-sampling and weighted loss. Still, I got a very low precision score on the activity Start Hesitation and Walking :(  For example, I got an average score 0.272, and 0.077,0.687,0.052 for  Start Hesitation, Turn, and Walking.</p>\n<p>2) is this a classification problem or a regression problem?</p>\n<p>I see many notebooks model this as a regression problem. Is there an issue if we model it as a classification problem?</p>\n<p>Thank you</p>",
      "rawMarkdown": "Hi,\n\nI have some questions about this competition, and I hope there is anyone who could help with this :)\n\n1) How should we address the imbalance dataset issue in this competition? \n\nI tried a couple of approaches like data up-sampling and weighted loss. Still, I got a very low precision score on the activity Start Hesitation and Walking :(  For example, I got an average score 0.272, and 0.077,0.687,0.052 for  Start Hesitation, Turn, and Walking.\n\n2) is this a classification problem or a regression problem?\n\nI see many notebooks model this as a regression problem. Is there an issue if we model it as a classification problem?\n\nThank you",
      "votes": null
    },
    {
      "id": "2256889",
      "postDate": "05/12/2023 20:01:45",
      "content": "<p>Here's my thoughts:<br>\n1) You have two ways to solve this issue - either from data side (undersampling/oversampling), or from model side (using focal loss, class weights, etc)<br>\n2) This is a multiclass classification problem (4 classes - no FOG event, Start Hesitation, Turn or Walking). I'm not sure why many popular notebooks use regression.</p>",
      "rawMarkdown": "Here's my thoughts:\n1) You have two ways to solve this issue - either from data side (undersampling/oversampling), or from model side (using focal loss, class weights, etc)\n2) This is a multiclass classification problem (4 classes - no FOG event, Start Hesitation, Turn or Walking). I'm not sure why many popular notebooks use regression.",
      "votes": null
    },
    {
      "id": "2259386",
      "postDate": "05/15/2023 00:42:11",
      "content": "<p>i try to model this as multiclass classification problem, but the result is not so good, i think is because of the metric problem.</p>",
      "rawMarkdown": "i try to model this as multiclass classification problem, but the result is not so good, i think is because of the metric problem.",
      "votes": null
    },
    {
      "id": "2264917",
      "postDate": "05/18/2023 20:49:45",
      "content": "<p>I'm going to try a DeepCNN-LSTM approach based on this paper: <a href=\"https://ieeexplore.ieee.org/abstract/document/8965723\" target=\"_blank\">https://ieeexplore.ieee.org/abstract/document/8965723</a></p>",
      "rawMarkdown": "I'm going to try a DeepCNN-LSTM approach based on this paper: https://ieeexplore.ieee.org/abstract/document/8965723",
      "votes": null
    },
    {
      "id": "2264918",
      "postDate": "05/18/2023 20:50:08",
      "content": "<p>Here's a helpful Keras implementation <a href=\"https://keras.io/examples/timeseries/timeseries_transformer_classification/\" target=\"_blank\">https://keras.io/examples/timeseries/timeseries_transformer_classification/</a></p>",
      "rawMarkdown": "Here's a helpful Keras implementation https://keras.io/examples/timeseries/timeseries_transformer_classification/",
      "votes": null
    },
    {
      "id": "2288512",
      "postDate": "06/05/2023 12:36:13",
      "content": "<p>why did you consider \"no FOG event\" as one of the classes for this multiclass classification problem when in the data description it is stated that \"Your objective is to <strong>detect the start and stop of each freezing episode and the occurrence</strong> in these series of three types of freezing of gait events: Start Hesitation, Turn, and Walking.\"?</p>",
      "rawMarkdown": "why did you consider \"no FOG event\" as one of the classes for this multiclass classification problem when in the data description it is stated that \"Your objective is to **detect the start and stop of each freezing episode and the occurrence** in these series of three types of freezing of gait events: Start Hesitation, Turn, and Walking.\"?",
      "votes": null
    },
    {
      "id": "2288553",
      "postDate": "06/05/2023 13:15:02",
      "content": "<p>Because there's a plenty of cases where no FOG event was detected (meaning that StartHesitation, Turn, and Walking are all equal to 0 at the same time).</p>",
      "rawMarkdown": "Because there's a plenty of cases where no FOG event was detected (meaning that StartHesitation, Turn, and Walking are all equal to 0 at the same time).",
      "votes": null
    },
    {
      "id": "2288834",
      "postDate": "06/05/2023 17:49:36",
      "content": "<p>if you used event data as base reference, there are 411 tdcsfog ids and 91 defog ids with values of 1 on either StartHesitation, Turn or Walking.</p>\n<p>411 --&gt; tdcsfog<br>\n    group_1: 32  --&gt; start hesitation, turn and walking<br>\n    group_2: 52  --&gt; start hesitation and turn<br>\n    group_3: 2   --&gt; start hesitation and walking<br>\n    group_4: 28  --&gt; turn and walking<br>\n    group_5: 7   --&gt; start hesitation<br>\n    group_6: 284 --&gt; turn<br>\n    group_7: 6   --&gt; walking</p>\n<p>91 --&gt; defog<br>\n    group_1: 5   --&gt; start hesitation, turn and walking<br>\n    group_2: 2   --&gt; start hesitation and turn<br>\n    group_3: 0   --&gt; start hesitation and walking<br>\n    group_4: 49  --&gt; turn and walking<br>\n    group_5: 0   --&gt; start hesitation<br>\n    group_6: 34  --&gt; turn<br>\n    group_7: 1   --&gt; walking</p>",
      "rawMarkdown": "if you used event data as base reference, there are 411 tdcsfog ids and 91 defog ids with values of 1 on either StartHesitation, Turn or Walking.\n\n411 --> tdcsfog\n\tgroup_1: 32  --> start hesitation, turn and walking\n\tgroup_2: 52  --> start hesitation and turn\n\tgroup_3: 2   --> start hesitation and walking\n\tgroup_4: 28  --> turn and walking\n\tgroup_5: 7   --> start hesitation\n\tgroup_6: 284 --> turn\n\tgroup_7: 6   --> walking\n\n91 --> defog\n\tgroup_1: 5   --> start hesitation, turn and walking\n\tgroup_2: 2   --> start hesitation and turn\n\tgroup_3: 0   --> start hesitation and walking\n\tgroup_4: 49  --> turn and walking\n\tgroup_5: 0   --> start hesitation\n\tgroup_6: 34  --> turn\n\tgroup_7: 1   --> walking",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2256889,
      "author_name": "atamazian",
      "author_url": "",
      "post_date": "05/12/2023 20:01:45",
      "content": "<p>Here's my thoughts:<br>\n1) You have two ways to solve this issue - either from data side (undersampling/oversampling), or from model side (using focal loss, class weights, etc)<br>\n2) This is a multiclass classification problem (4 classes - no FOG event, Start Hesitation, Turn or Walking). I'm not sure why many popular notebooks use regression.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2259386,
          "author_name": "xianzwaikato",
          "author_url": "",
          "post_date": "05/15/2023 00:42:11",
          "content": "<p>i try to model this as multiclass classification problem, but the result is not so good, i think is because of the metric problem.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 2288512,
          "author_name": "projdev",
          "author_url": "",
          "post_date": "06/05/2023 12:36:13",
          "content": "<p>why did you consider \"no FOG event\" as one of the classes for this multiclass classification problem when in the data description it is stated that \"Your objective is to <strong>detect the start and stop of each freezing episode and the occurrence</strong> in these series of three types of freezing of gait events: Start Hesitation, Turn, and Walking.\"?</p>",
          "votes": null,
          "replies": [
            {
              "id": 2288553,
              "author_name": "atamazian",
              "author_url": "",
              "post_date": "06/05/2023 13:15:02",
              "content": "<p>Because there's a plenty of cases where no FOG event was detected (meaning that StartHesitation, Turn, and Walking are all equal to 0 at the same time).</p>",
              "votes": null,
              "replies": [
                {
                  "id": 2288834,
                  "author_name": "projdev",
                  "author_url": "",
                  "post_date": "06/05/2023 17:49:36",
                  "content": "<p>if you used event data as base reference, there are 411 tdcsfog ids and 91 defog ids with values of 1 on either StartHesitation, Turn or Walking.</p>\n<p>411 --&gt; tdcsfog<br>\n    group_1: 32  --&gt; start hesitation, turn and walking<br>\n    group_2: 52  --&gt; start hesitation and turn<br>\n    group_3: 2   --&gt; start hesitation and walking<br>\n    group_4: 28  --&gt; turn and walking<br>\n    group_5: 7   --&gt; start hesitation<br>\n    group_6: 284 --&gt; turn<br>\n    group_7: 6   --&gt; walking</p>\n<p>91 --&gt; defog<br>\n    group_1: 5   --&gt; start hesitation, turn and walking<br>\n    group_2: 2   --&gt; start hesitation and turn<br>\n    group_3: 0   --&gt; start hesitation and walking<br>\n    group_4: 49  --&gt; turn and walking<br>\n    group_5: 0   --&gt; start hesitation<br>\n    group_6: 34  --&gt; turn<br>\n    group_7: 1   --&gt; walking</p>",
                  "votes": null,
                  "replies": []
                }
              ]
            }
          ]
        }
      ]
    },
    {
      "id": 2264917,
      "author_name": "leewhieldon",
      "author_url": "",
      "post_date": "05/18/2023 20:49:45",
      "content": "<p>I'm going to try a DeepCNN-LSTM approach based on this paper: <a href=\"https://ieeexplore.ieee.org/abstract/document/8965723\" target=\"_blank\">https://ieeexplore.ieee.org/abstract/document/8965723</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 2264918,
          "author_name": "leewhieldon",
          "author_url": "",
          "post_date": "05/18/2023 20:50:08",
          "content": "<p>Here's a helpful Keras implementation <a href=\"https://keras.io/examples/timeseries/timeseries_transformer_classification/\" target=\"_blank\">https://keras.io/examples/timeseries/timeseries_transformer_classification/</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2240950": "Hi,\n\nI have some questions about this competition, and I hope there is anyone who could help with this :)\n\n1) How should we address the imbalance dataset issue in this competition? \n\nI tried a couple of approaches like data up-sampling and weighted loss. Still, I got a very low precision score on the activity Start Hesitation and Walking :(  For example, I got an average score 0.272, and 0.077,0.687,0.052 for  Start Hesitation, Turn, and Walking.\n\n2) is this a classification problem or a regression problem?\n\nI see many notebooks model this as a regression problem. Is there an issue if we model it as a classification problem?\n\nThank you",
    "2256889": "Here's my thoughts:\n1) You have two ways to solve this issue - either from data side (undersampling/oversampling), or from model side (using focal loss, class weights, etc)\n2) This is a multiclass classification problem (4 classes - no FOG event, Start Hesitation, Turn or Walking). I'm not sure why many popular notebooks use regression.",
    "2259386": "i try to model this as multiclass classification problem, but the result is not so good, i think is because of the metric problem.",
    "2264917": "I'm going to try a DeepCNN-LSTM approach based on this paper: https://ieeexplore.ieee.org/abstract/document/8965723",
    "2264918": "Here's a helpful Keras implementation https://keras.io/examples/timeseries/timeseries_transformer_classification/",
    "2288512": "why did you consider \"no FOG event\" as one of the classes for this multiclass classification problem when in the data description it is stated that \"Your objective is to **detect the start and stop of each freezing episode and the occurrence** in these series of three types of freezing of gait events: Start Hesitation, Turn, and Walking.\"?",
    "2288553": "Because there's a plenty of cases where no FOG event was detected (meaning that StartHesitation, Turn, and Walking are all equal to 0 at the same time).",
    "2288834": "if you used event data as base reference, there are 411 tdcsfog ids and 91 defog ids with values of 1 on either StartHesitation, Turn or Walking.\n\n411 --> tdcsfog\n\tgroup_1: 32  --> start hesitation, turn and walking\n\tgroup_2: 52  --> start hesitation and turn\n\tgroup_3: 2   --> start hesitation and walking\n\tgroup_4: 28  --> turn and walking\n\tgroup_5: 7   --> start hesitation\n\tgroup_6: 284 --> turn\n\tgroup_7: 6   --> walking\n\n91 --> defog\n\tgroup_1: 5   --> start hesitation, turn and walking\n\tgroup_2: 2   --> start hesitation and turn\n\tgroup_3: 0   --> start hesitation and walking\n\tgroup_4: 49  --> turn and walking\n\tgroup_5: 0   --> start hesitation\n\tgroup_6: 34  --> turn\n\tgroup_7: 1   --> walking"
  },
  "source": "meta"
}