{
  "id": 544316,
  "title": "1st place: What happened to our team during the last 2 days?",
  "url": "/competitions/ariel-data-challenge-2024/discussion/544316",
  "author_name": "c-number",
  "post_date": "2024-11-04T13:30:00.423000",
  "votes": 23,
  "comment_count": 3,
  "views": 0,
  "content": "<p><a href=\"https://www.kaggle.com/competitions/ariel-data-challenge-2024/discussion/544317\" target=\"_blank\">link to solution</a></p>\n<p>Some may have notice that we managed to climb up the LB very quick during the last 2 days, from 0.694 to 0.733.</p>\n<h1>So, what happened?</h1>\n<p>The turning point was that we found out that removing the hot pixel processing drastically enhanced the LB score by +0.015 despite not having much effect on the training score, and not only that, but the removal made multiple processes such as two-stage fitting, AutoEncoder, and NMF that didn't work on the LB score before magically work.</p>\n<p>Also, we were lucky to be able to find the correct adjustment of sigma, etc., with our final submission because we couldn't have won first place with our second-to-last submission.</p>\n<h1>Submissions</h1>\n<table>\n<thead>\n<tr>\n<th>Submission</th>\n<th>Remaining Hours</th>\n<th>Public</th>\n<th>Public Gain Compared to Past Best Sub</th>\n<th>Private</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Highest score before the last 2 days</td>\n<td>48</td>\n<td>0.6948251</td>\n<td></td>\n<td>0.7086383</td>\n</tr>\n<tr>\n<td>Remove Hot Pixel Processing</td>\n<td>40</td>\n<td>0.7104020</td>\n<td>+0.0155769</td>\n<td>0.7206170</td>\n</tr>\n<tr>\n<td>Remove Foreground Processing, with *1.007</td>\n<td>37</td>\n<td>0.7098685</td>\n<td>-0.0005335</td>\n<td>0.7159903</td>\n</tr>\n<tr>\n<td>Gain Drift Fitting to Two-Stage</td>\n<td>24</td>\n<td>0.7208129</td>\n<td>+0.0104109</td>\n<td>0.7303065</td>\n</tr>\n<tr>\n<td>Add AutoEncoder</td>\n<td>18</td>\n<td>0.7258255</td>\n<td>+0.0050126</td>\n<td>0.7352213</td>\n</tr>\n<tr>\n<td>Add NMF</td>\n<td>11</td>\n<td>0.7281795</td>\n<td>+0.0023540</td>\n<td>0.7390831</td>\n</tr>\n<tr>\n<td>Final Adjustment of Sigma, etc.</td>\n<td>6</td>\n<td>0.7330321</td>\n<td>+0.0048526</td>\n<td>0.7420624</td>\n</tr>\n</tbody>\n</table>\n<h1>Score Timelapse</h1>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6624777%2F1c91e9bf05c6cc916a84efbcd0d11718%2Fchart.png?generation=1730726956265209&amp;alt=media\" alt=\"\"></p>",
  "messages": [
    {
      "id": 3036335,
      "postDate": "2024-11-04T13:30:00.423Z",
      "content": "<p><a href=\"https://www.kaggle.com/competitions/ariel-data-challenge-2024/discussion/544317\" target=\"_blank\">link to solution</a></p>\n<p>Some may have notice that we managed to climb up the LB very quick during the last 2 days, from 0.694 to 0.733.</p>\n<h1>So, what happened?</h1>\n<p>The turning point was that we found out that removing the hot pixel processing drastically enhanced the LB score by +0.015 despite not having much effect on the training score, and not only that, but the removal made multiple processes such as two-stage fitting, AutoEncoder, and NMF that didn't work on the LB score before magically work.</p>\n<p>Also, we were lucky to be able to find the correct adjustment of sigma, etc., with our final submission because we couldn't have won first place with our second-to-last submission.</p>\n<h1>Submissions</h1>\n<table>\n<thead>\n<tr>\n<th>Submission</th>\n<th>Remaining Hours</th>\n<th>Public</th>\n<th>Public Gain Compared to Past Best Sub</th>\n<th>Private</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Highest score before the last 2 days</td>\n<td>48</td>\n<td>0.6948251</td>\n<td></td>\n<td>0.7086383</td>\n</tr>\n<tr>\n<td>Remove Hot Pixel Processing</td>\n<td>40</td>\n<td>0.7104020</td>\n<td>+0.0155769</td>\n<td>0.7206170</td>\n</tr>\n<tr>\n<td>Remove Foreground Processing, with *1.007</td>\n<td>37</td>\n<td>0.7098685</td>\n<td>-0.0005335</td>\n<td>0.7159903</td>\n</tr>\n<tr>\n<td>Gain Drift Fitting to Two-Stage</td>\n<td>24</td>\n<td>0.7208129</td>\n<td>+0.0104109</td>\n<td>0.7303065</td>\n</tr>\n<tr>\n<td>Add AutoEncoder</td>\n<td>18</td>\n<td>0.7258255</td>\n<td>+0.0050126</td>\n<td>0.7352213</td>\n</tr>\n<tr>\n<td>Add NMF</td>\n<td>11</td>\n<td>0.7281795</td>\n<td>+0.0023540</td>\n<td>0.7390831</td>\n</tr>\n<tr>\n<td>Final Adjustment of Sigma, etc.</td>\n<td>6</td>\n<td>0.7330321</td>\n<td>+0.0048526</td>\n<td>0.7420624</td>\n</tr>\n</tbody>\n</table>\n<h1>Score Timelapse</h1>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6624777%2F1c91e9bf05c6cc916a84efbcd0d11718%2Fchart.png?generation=1730726956265209&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "[link to solution](https://www.kaggle.com/competitions/ariel-data-challenge-2024/discussion/544317)\n\nSome may have notice that we managed to climb up the LB very quick during the last 2 days, from 0.694 to 0.733.\n\n# So, what happened?\n\nThe turning point was that we found out that removing the hot pixel processing drastically enhanced the LB score by +0.015 despite not having much effect on the training score, and not only that, but the removal made multiple processes such as two-stage fitting, AutoEncoder, and NMF that didn't work on the LB score before magically work.\n\nAlso, we were lucky to be able to find the correct adjustment of sigma, etc., with our final submission because we couldn't have won first place with our second-to-last submission.\n\n\n# Submissions\n\n| Submission                                 | Remaining Hours | Public     | Public Gain Compared to Past Best Sub | Private    |\n|--------------------------------------------|-----------------|------------|---------------------------------------|------------|\n| Highest score before the last 2 days       | 48              | 0.6948251  |                                       | 0.7086383  |\n| Remove Hot Pixel Processing                | 40              | 0.7104020  | +0.0155769                            | 0.7206170  |\n| Remove Foreground Processing, with *1.007  | 37              | 0.7098685  | -0.0005335                            | 0.7159903  |\n| Gain Drift Fitting to Two-Stage            | 24              | 0.7208129  | +0.0104109                            | 0.7303065  |\n| Add AutoEncoder                            | 18              | 0.7258255  | +0.0050126                            | 0.7352213  |\n| Add NMF                                    | 11              | 0.7281795  | +0.0023540                            | 0.7390831  |\n| Final Adjustment of Sigma, etc.            | 6               | 0.7330321  | +0.0048526                            | 0.7420624  |\n\n# Score Timelapse\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6624777%2F1c91e9bf05c6cc916a84efbcd0d11718%2Fchart.png?generation=1730726956265209&alt=media)",
      "votes": 23
    },
    {
      "id": 3036365,
      "postDate": "2024-11-04T13:45:50.930Z",
      "content": "<p>I was checking the leaderboard pretty often that day and was convinced you were out of submission on your penultimate 0.725 result, so was pretty shocked later…did you make a submission just before the deadline the day before?</p>",
      "rawMarkdown": "I was checking the leaderboard pretty often that day and was convinced you were out of submission on your penultimate 0.725 result, so was pretty shocked later...did you make a submission just before the deadline the day before?",
      "votes": 1,
      "replies": [
        {
          "id": 3036379,
          "postDate": "2024-11-04T14:09:18.983Z",
          "content": "<p>The 4th-to-last submission was made just 1 minuite before the day change to literally add some last minuite adjusments, so that might be the reason.<br>\nYou can hover over the submission time on the LB (4d in the image below) to obtain the exact submission time, just to let you know.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6624777%2Ffec3f52c0e6c52a5bc802c49fb187828%2FScreenshot%20from%202024-11-04%2023-06-21.png?generation=1730729190509716&amp;alt=media\" alt=\"\"></p>",
          "rawMarkdown": "The 4th-to-last submission was made just 1 minuite before the day change to literally add some last minuite adjusments, so that might be the reason.\nYou can hover over the submission time on the LB (4d in the image below) to obtain the exact submission time, just to let you know.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6624777%2Ffec3f52c0e6c52a5bc802c49fb187828%2FScreenshot%20from%202024-11-04%2023-06-21.png?generation=1730729190509716&alt=media)",
          "votes": 4,
          "replies": [
            {
              "id": 3036431,
              "postDate": "2024-11-04T14:43:28.357Z",
              "content": "<p>That's what I did, but I didn't look carefully and assumed you submitted one just <em>after</em> the deadline…</p>",
              "rawMarkdown": "That's what I did, but I didn't look carefully and assumed you submitted one just *after* the deadline..."
            }
          ]
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 3036365,
      "author_name": "Jeroen Cottaar",
      "author_url": "",
      "post_date": "2024-11-04T13:45:50.930000",
      "content": "<p>I was checking the leaderboard pretty often that day and was convinced you were out of submission on your penultimate 0.725 result, so was pretty shocked later…did you make a submission just before the deadline the day before?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3036379,
          "author_name": "c-number",
          "author_url": "",
          "post_date": "2024-11-04T14:09:18.983000",
          "content": "<p>The 4th-to-last submission was made just 1 minuite before the day change to literally add some last minuite adjusments, so that might be the reason.<br>\nYou can hover over the submission time on the LB (4d in the image below) to obtain the exact submission time, just to let you know.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6624777%2Ffec3f52c0e6c52a5bc802c49fb187828%2FScreenshot%20from%202024-11-04%2023-06-21.png?generation=1730729190509716&amp;alt=media\" alt=\"\"></p>",
          "votes": 4,
          "replies": [
            {
              "id": 3036431,
              "author_name": "Jeroen Cottaar",
              "author_url": "",
              "post_date": "2024-11-04T14:43:28.357000",
              "content": "<p>That's what I did, but I didn't look carefully and assumed you submitted one just <em>after</em> the deadline…</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3036335": "[link to solution](https://www.kaggle.com/competitions/ariel-data-challenge-2024/discussion/544317)\n\nSome may have notice that we managed to climb up the LB very quick during the last 2 days, from 0.694 to 0.733.\n\n# So, what happened?\n\nThe turning point was that we found out that removing the hot pixel processing drastically enhanced the LB score by +0.015 despite not having much effect on the training score, and not only that, but the removal made multiple processes such as two-stage fitting, AutoEncoder, and NMF that didn't work on the LB score before magically work.\n\nAlso, we were lucky to be able to find the correct adjustment of sigma, etc., with our final submission because we couldn't have won first place with our second-to-last submission.\n\n\n# Submissions\n\n| Submission                                 | Remaining Hours | Public     | Public Gain Compared to Past Best Sub | Private    |\n|--------------------------------------------|-----------------|------------|---------------------------------------|------------|\n| Highest score before the last 2 days       | 48              | 0.6948251  |                                       | 0.7086383  |\n| Remove Hot Pixel Processing                | 40              | 0.7104020  | +0.0155769                            | 0.7206170  |\n| Remove Foreground Processing, with *1.007  | 37              | 0.7098685  | -0.0005335                            | 0.7159903  |\n| Gain Drift Fitting to Two-Stage            | 24              | 0.7208129  | +0.0104109                            | 0.7303065  |\n| Add AutoEncoder                            | 18              | 0.7258255  | +0.0050126                            | 0.7352213  |\n| Add NMF                                    | 11              | 0.7281795  | +0.0023540                            | 0.7390831  |\n| Final Adjustment of Sigma, etc.            | 6               | 0.7330321  | +0.0048526                            | 0.7420624  |\n\n# Score Timelapse\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6624777%2F1c91e9bf05c6cc916a84efbcd0d11718%2Fchart.png?generation=1730726956265209&alt=media)",
    "3036365": "I was checking the leaderboard pretty often that day and was convinced you were out of submission on your penultimate 0.725 result, so was pretty shocked later...did you make a submission just before the deadline the day before?"
  }
}