{
  "id": 63320,
  "title": "Pretty pictures: \"Robin, get the Bat-Booster!\"",
  "url": "/competitions/trackml-particle-identification/discussion/63320",
  "author_name": "Edwin Steiner",
  "post_date": "2018-08-14T20:20:12.297000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Here's a nice picture I generated, though I ended up not using the model behind it.</p>\n\n<p>The graph shows azimuthal displacements of positive particles from perfect helix predictions in outer cylinders and outer caps. The vertical unit is azimuthal displacement as a multiple of the measurement error. The horizontal unit is z-coordinate in mm.</p>\n\n<p>The graph shows to which degree an xgboost regressor which I trained could predict the displacements. (I did not use this model in the end.)</p>\n\n<p>I cannot explain the whole graph, but briefly:</p>\n\n<ul>\n<li>red and blue plots and green histograms show the distribution of the displacements without correction</li>\n<li>the yellow dashed line shows the RMS displacement before prediction</li>\n<li>yellow dots show binned standard deviation of the displacements (i.e. the random RMS if only the z-dependent mean displacement is considered systematic)</li>\n<li>the orange dashed line shows the RMS deviation after subtracting the prediction</li>\n<li>the gray \"Batman symbol\" line shows the typical azimuthal distance to the nearest background hit in the same units. (Note that it is much larger than the RMS displacement, which explains why implementing this model had hardly any effect on the score.)</li>\n</ul>\n\n<p>The fact that the orange line is lower than the yellow line and even  lower than the yellow dots means that the xgboost regressor did a decent job of predicting the displacements, better than a pure z-based binned mean would have done (yellow dots).</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/370433/10079/bat_booster_dephi.png\" alt=\"plot of azimuthal displacements\"></p>",
  "messages": [
    {
      "id": 370433,
      "postDate": "2018-08-14T20:20:12.297Z",
      "content": "<p>Here's a nice picture I generated, though I ended up not using the model behind it.</p>\n\n<p>The graph shows azimuthal displacements of positive particles from perfect helix predictions in outer cylinders and outer caps. The vertical unit is azimuthal displacement as a multiple of the measurement error. The horizontal unit is z-coordinate in mm.</p>\n\n<p>The graph shows to which degree an xgboost regressor which I trained could predict the displacements. (I did not use this model in the end.)</p>\n\n<p>I cannot explain the whole graph, but briefly:</p>\n\n<ul>\n<li>red and blue plots and green histograms show the distribution of the displacements without correction</li>\n<li>the yellow dashed line shows the RMS displacement before prediction</li>\n<li>yellow dots show binned standard deviation of the displacements (i.e. the random RMS if only the z-dependent mean displacement is considered systematic)</li>\n<li>the orange dashed line shows the RMS deviation after subtracting the prediction</li>\n<li>the gray \"Batman symbol\" line shows the typical azimuthal distance to the nearest background hit in the same units. (Note that it is much larger than the RMS displacement, which explains why implementing this model had hardly any effect on the score.)</li>\n</ul>\n\n<p>The fact that the orange line is lower than the yellow line and even  lower than the yellow dots means that the xgboost regressor did a decent job of predicting the displacements, better than a pure z-based binned mean would have done (yellow dots).</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/370433/10079/bat_booster_dephi.png\" alt=\"plot of azimuthal displacements\"></p>",
      "rawMarkdown": "Here's a nice picture I generated, though I ended up not using the model behind it.\n\nThe graph shows azimuthal displacements of positive particles from perfect helix predictions in outer cylinders and outer caps. The vertical unit is azimuthal displacement as a multiple of the measurement error. The horizontal unit is z-coordinate in mm.\n\nThe graph shows to which degree an xgboost regressor which I trained could predict the displacements. (I did not use this model in the end.)\n\nI cannot explain the whole graph, but briefly:\n\n - red and blue plots and green histograms show the distribution of the displacements without correction\n - the yellow dashed line shows the RMS displacement before prediction\n - yellow dots show binned standard deviation of the displacements (i.e. the random RMS if only the z-dependent mean displacement is considered systematic)\n - the orange dashed line shows the RMS deviation after subtracting the prediction\n - the gray \"Batman symbol\" line shows the typical azimuthal distance to the nearest background hit in the same units. (Note that it is much larger than the RMS displacement, which explains why implementing this model had hardly any effect on the score.)\n\nThe fact that the orange line is lower than the yellow line and even  lower than the yellow dots means that the xgboost regressor did a decent job of predicting the displacements, better than a pure z-based binned mean would have done (yellow dots).\n\n![plot of azimuthal displacements][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/370433/10079/bat_booster_dephi.png",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "370433": "Here's a nice picture I generated, though I ended up not using the model behind it.\n\nThe graph shows azimuthal displacements of positive particles from perfect helix predictions in outer cylinders and outer caps. The vertical unit is azimuthal displacement as a multiple of the measurement error. The horizontal unit is z-coordinate in mm.\n\nThe graph shows to which degree an xgboost regressor which I trained could predict the displacements. (I did not use this model in the end.)\n\nI cannot explain the whole graph, but briefly:\n\n - red and blue plots and green histograms show the distribution of the displacements without correction\n - the yellow dashed line shows the RMS displacement before prediction\n - yellow dots show binned standard deviation of the displacements (i.e. the random RMS if only the z-dependent mean displacement is considered systematic)\n - the orange dashed line shows the RMS deviation after subtracting the prediction\n - the gray \"Batman symbol\" line shows the typical azimuthal distance to the nearest background hit in the same units. (Note that it is much larger than the RMS displacement, which explains why implementing this model had hardly any effect on the score.)\n\nThe fact that the orange line is lower than the yellow line and even  lower than the yellow dots means that the xgboost regressor did a decent job of predicting the displacements, better than a pure z-based binned mean would have done (yellow dots).\n\n![plot of azimuthal displacements][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/370433/10079/bat_booster_dephi.png"
  }
}