{
  "id": 216924,
  "title": "Questions and Speculations on How To Handle the Negative Class",
  "url": "/competitions/hpa-single-cell-image-classification/discussion/216924",
  "author_name": "Darien Schettler",
  "post_date": "2021-02-04T14:31:01.913000",
  "votes": 11,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi there.</p>\n<p>I have a couple of questions/speculations around the negative class I wanted to clear-up/air.</p>\n<p><br></p>\n<p><strong>Questions</strong></p>\n<hr>\n<ol>\n<li>Will every cell in an image have a label? If so, cells with the protein-of-interest showing the correct pattern will be labelled accordingly, but cells with no protein-of-interest showing, or a noisy signal, should be labelled as negative. Is this assumption correct?</li>\n<li>Since a cell can have multiple labels, is it safe to assume we will <strong>never</strong> have a cell labelled as <strong><code>Negative</code> AND <code>some-other-label</code></strong>? </li>\n</ol>\n<hr>\n<p><br></p>\n<p><strong>Speculations</strong></p>\n<hr>\n<p><em>If I assume the above questions are answered as per my current understanding…</em></p>\n<ol>\n<li>We will be able to identify protein patterns/coloring in the green-channel for a given label, which we can then use to re-map an image-level class to negative for cells not presenting that required pattern. All the other cells will either get the image-level label (single-class image) or some distribution of the image-level labels (multi-class image).</li>\n<li>There will be quite a lot of negative class cells (much higher representation than at the image level) due to a portion of cells in every image not presenting the desired protein-of-interest patterns, or displaying a noisy signal.</li>\n</ol>\n<hr>\n<p><br></p>\n<p>Thanks for taking the time to read. Please let me know what your thoughts are and if my assumptions and speculations make sense!</p>",
  "messages": [
    {
      "id": 1186010,
      "postDate": "2021-02-04T14:31:01.913Z",
      "content": "<p>Hi there.</p>\n<p>I have a couple of questions/speculations around the negative class I wanted to clear-up/air.</p>\n<p><br></p>\n<p><strong>Questions</strong></p>\n<hr>\n<ol>\n<li>Will every cell in an image have a label? If so, cells with the protein-of-interest showing the correct pattern will be labelled accordingly, but cells with no protein-of-interest showing, or a noisy signal, should be labelled as negative. Is this assumption correct?</li>\n<li>Since a cell can have multiple labels, is it safe to assume we will <strong>never</strong> have a cell labelled as <strong><code>Negative</code> AND <code>some-other-label</code></strong>? </li>\n</ol>\n<hr>\n<p><br></p>\n<p><strong>Speculations</strong></p>\n<hr>\n<p><em>If I assume the above questions are answered as per my current understanding…</em></p>\n<ol>\n<li>We will be able to identify protein patterns/coloring in the green-channel for a given label, which we can then use to re-map an image-level class to negative for cells not presenting that required pattern. All the other cells will either get the image-level label (single-class image) or some distribution of the image-level labels (multi-class image).</li>\n<li>There will be quite a lot of negative class cells (much higher representation than at the image level) due to a portion of cells in every image not presenting the desired protein-of-interest patterns, or displaying a noisy signal.</li>\n</ol>\n<hr>\n<p><br></p>\n<p>Thanks for taking the time to read. Please let me know what your thoughts are and if my assumptions and speculations make sense!</p>",
      "rawMarkdown": "Hi there.\n\nI have a couple of questions/speculations around the negative class I wanted to clear-up/air.\n\n<br>\n\n**Questions**\n\n---\n\n1. Will every cell in an image have a label? If so, cells with the protein-of-interest showing the correct pattern will be labelled accordingly, but cells with no protein-of-interest showing, or a noisy signal, should be labelled as negative. Is this assumption correct?\n2. Since a cell can have multiple labels, is it safe to assume we will **never** have a cell labelled as **`Negative` AND `some-other-label`**? \n\n---\n\n<br>\n\n**Speculations**\n\n---\n\n*If I assume the above questions are answered as per my current understanding...*\n\n1.  We will be able to identify protein patterns/coloring in the green-channel for a given label, which we can then use to re-map an image-level class to negative for cells not presenting that required pattern. All the other cells will either get the image-level label (single-class image) or some distribution of the image-level labels (multi-class image).\n2. There will be quite a lot of negative class cells (much higher representation than at the image level) due to a portion of cells in every image not presenting the desired protein-of-interest patterns, or displaying a noisy signal.\n\n---\n\n<br>\n\nThanks for taking the time to read. Please let me know what your thoughts are and if my assumptions and speculations make sense!",
      "votes": 11
    },
    {
      "id": 1186174,
      "postDate": "2021-02-04T16:14:28.313Z",
      "content": "<p>I agree the <code>Negative</code> label is a bit unclear, and I too hope for answers of the above.</p>\n<p>Also, in the training data, can we consider this category as a \"garbage term\" or a \"none of the above\". If so, wouldn't it be equivalent to a category vector consisting only of zeros (or False or Null etc). Or is there some subtle distinction I don't see?</p>",
      "rawMarkdown": "I agree the `Negative` label is a bit unclear, and I too hope for answers of the above.\n\nAlso, in the training data, can we consider this category as a \"garbage term\" or a \"none of the above\". If so, wouldn't it be equivalent to a category vector consisting only of zeros (or False or Null etc). Or is there some subtle distinction I don't see?",
      "votes": 1,
      "replies": [
        {
          "id": 1187656,
          "postDate": "2021-02-05T15:58:05.710Z",
          "content": "<p>Yes you are correct, it is a \"none of the above\" label, and can be treated as such. For practical scoring reasons on the Kaggle platform we're using the format where 'Negative' has a label of its own (18).</p>",
          "rawMarkdown": "Yes you are correct, it is a \"none of the above\" label, and can be treated as such. For practical scoring reasons on the Kaggle platform we're using the format where 'Negative' has a label of its own (18).",
          "votes": 2
        }
      ]
    },
    {
      "id": 1187646,
      "postDate": "2021-02-05T15:52:22.960Z",
      "content": "<p>Hi,</p>\n<p>Very good questions that we would gladly clarify.</p>\n<p>For the <strong>questions</strong>, you are completely correct. </p>\n<ol>\n<li>Every cell in an image will have one or multiple labels. If there is no specific green pattern (noisy or no signal) it should be labeled as negative. Note that border cells are only included as ground truth when there is enough information to decide on the label (as described in the <a href=\"https://www.kaggle.com/c/hpa-single-cell-image-classification/data\" target=\"_blank\">data</a> tab, and the <a href=\"https://www.kaggle.com/lnhtrang/single-cell-patterns\" target=\"_blank\">single cell pattern notebook</a>), i.e. when the majority of the cell area is represented in the image.</li>\n<li>A cell can have multiple labels, but will <strong>never</strong> have <code>Negative</code> AND <code>some-other-label</code>.</li>\n</ol>\n<p>For the <strong>speculations</strong> you are mainly correct, but I would like to point out that there may also be images where a single cell have a label that was neglected at the image level (for example a weak cytosolic pattern).</p>",
      "rawMarkdown": "Hi,\n\nVery good questions that we would gladly clarify.\n\nFor the **questions**, you are completely correct. \n1. Every cell in an image will have one or multiple labels. If there is no specific green pattern (noisy or no signal) it should be labeled as negative. Note that border cells are only included as ground truth when there is enough information to decide on the label (as described in the [data](https://www.kaggle.com/c/hpa-single-cell-image-classification/data) tab, and the [single cell pattern notebook](https://www.kaggle.com/lnhtrang/single-cell-patterns)), i.e. when the majority of the cell area is represented in the image.\n2. A cell can have multiple labels, but will **never** have `Negative` AND `some-other-label`.\n\nFor the **speculations** you are mainly correct, but I would like to point out that there may also be images where a single cell have a label that was neglected at the image level (for example a weak cytosolic pattern).",
      "votes": 2,
      "replies": [
        {
          "id": 1187702,
          "postDate": "2021-02-05T16:39:36.503Z",
          "content": "<p>Thank you for the response Emma. That helps clarify things!</p>",
          "rawMarkdown": "Thank you for the response Emma. That helps clarify things!"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1186174,
      "author_name": "AndersOhrn",
      "author_url": "",
      "post_date": "2021-02-04T16:14:28.313000",
      "content": "<p>I agree the <code>Negative</code> label is a bit unclear, and I too hope for answers of the above.</p>\n<p>Also, in the training data, can we consider this category as a \"garbage term\" or a \"none of the above\". If so, wouldn't it be equivalent to a category vector consisting only of zeros (or False or Null etc). Or is there some subtle distinction I don't see?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1187656,
          "author_name": "Emma Lundberg",
          "author_url": "",
          "post_date": "2021-02-05T15:58:05.710000",
          "content": "<p>Yes you are correct, it is a \"none of the above\" label, and can be treated as such. For practical scoring reasons on the Kaggle platform we're using the format where 'Negative' has a label of its own (18).</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 1187646,
      "author_name": "Emma Lundberg",
      "author_url": "",
      "post_date": "2021-02-05T15:52:22.960000",
      "content": "<p>Hi,</p>\n<p>Very good questions that we would gladly clarify.</p>\n<p>For the <strong>questions</strong>, you are completely correct. </p>\n<ol>\n<li>Every cell in an image will have one or multiple labels. If there is no specific green pattern (noisy or no signal) it should be labeled as negative. Note that border cells are only included as ground truth when there is enough information to decide on the label (as described in the <a href=\"https://www.kaggle.com/c/hpa-single-cell-image-classification/data\" target=\"_blank\">data</a> tab, and the <a href=\"https://www.kaggle.com/lnhtrang/single-cell-patterns\" target=\"_blank\">single cell pattern notebook</a>), i.e. when the majority of the cell area is represented in the image.</li>\n<li>A cell can have multiple labels, but will <strong>never</strong> have <code>Negative</code> AND <code>some-other-label</code>.</li>\n</ol>\n<p>For the <strong>speculations</strong> you are mainly correct, but I would like to point out that there may also be images where a single cell have a label that was neglected at the image level (for example a weak cytosolic pattern).</p>",
      "votes": 2,
      "replies": [
        {
          "id": 1187702,
          "author_name": "Darien Schettler",
          "author_url": "",
          "post_date": "2021-02-05T16:39:36.503000",
          "content": "<p>Thank you for the response Emma. That helps clarify things!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1186010": "Hi there.\n\nI have a couple of questions/speculations around the negative class I wanted to clear-up/air.\n\n<br>\n\n**Questions**\n\n---\n\n1. Will every cell in an image have a label? If so, cells with the protein-of-interest showing the correct pattern will be labelled accordingly, but cells with no protein-of-interest showing, or a noisy signal, should be labelled as negative. Is this assumption correct?\n2. Since a cell can have multiple labels, is it safe to assume we will **never** have a cell labelled as **`Negative` AND `some-other-label`**? \n\n---\n\n<br>\n\n**Speculations**\n\n---\n\n*If I assume the above questions are answered as per my current understanding...*\n\n1.  We will be able to identify protein patterns/coloring in the green-channel for a given label, which we can then use to re-map an image-level class to negative for cells not presenting that required pattern. All the other cells will either get the image-level label (single-class image) or some distribution of the image-level labels (multi-class image).\n2. There will be quite a lot of negative class cells (much higher representation than at the image level) due to a portion of cells in every image not presenting the desired protein-of-interest patterns, or displaying a noisy signal.\n\n---\n\n<br>\n\nThanks for taking the time to read. Please let me know what your thoughts are and if my assumptions and speculations make sense!",
    "1186174": "I agree the `Negative` label is a bit unclear, and I too hope for answers of the above.\n\nAlso, in the training data, can we consider this category as a \"garbage term\" or a \"none of the above\". If so, wouldn't it be equivalent to a category vector consisting only of zeros (or False or Null etc). Or is there some subtle distinction I don't see?",
    "1187646": "Hi,\n\nVery good questions that we would gladly clarify.\n\nFor the **questions**, you are completely correct. \n1. Every cell in an image will have one or multiple labels. If there is no specific green pattern (noisy or no signal) it should be labeled as negative. Note that border cells are only included as ground truth when there is enough information to decide on the label (as described in the [data](https://www.kaggle.com/c/hpa-single-cell-image-classification/data) tab, and the [single cell pattern notebook](https://www.kaggle.com/lnhtrang/single-cell-patterns)), i.e. when the majority of the cell area is represented in the image.\n2. A cell can have multiple labels, but will **never** have `Negative` AND `some-other-label`.\n\nFor the **speculations** you are mainly correct, but I would like to point out that there may also be images where a single cell have a label that was neglected at the image level (for example a weak cytosolic pattern)."
  }
}