{
  "id": 75049,
  "title": "Another mystery of 99",
  "url": "/competitions/PLAsTiCC-2018/discussion/75049",
  "author_name": "",
  "post_date": "2018-12-18T06:16:14.981804200Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I used Scirpus method to predict class 99 and sorted predictions and metadata by object_id. Then I plotted mjd differences for objects where prediction for class 99 is maximum and their indices and found something strange. \nFor cnn+gru model:\n<img src=\"https://i.imgur.com/FxRd4ue.png\" alt=\"enter image description here\"></p>\n\n<p>For <a href=\"https://www.kaggle.com/jimpsull/collaboratingwithkagglecommunity-1-037-lb\">CollaboratingWithKaggleCommunity (1.037 LB)</a> predictions there are few outliers in the same place:\n<img src=\"https://i.imgur.com/ScY98U4.png\" alt=\"enter image description here\">\nGraphs for mjd differences and object ids:\n<img src=\"https://i.imgur.com/kabFlbv.png\" alt=\"enter image description here\">\n<img src=\"https://i.imgur.com/tSna7fm.png\" alt=\"enter image description here\">\nCould it be some kind of flux leakage or it's just my mistake?</p>",
  "messages": [
    {
      "id": "440974",
      "postDate": "12/18/2018 06:16:14",
      "content": "<p>I used Scirpus method to predict class 99 and sorted predictions and metadata by object_id. Then I plotted mjd differences for objects where prediction for class 99 is maximum and their indices and found something strange. \nFor cnn+gru model:\n<img src=\"https://i.imgur.com/FxRd4ue.png\" alt=\"enter image description here\"></p>\n\n<p>For <a href=\"https://www.kaggle.com/jimpsull/collaboratingwithkagglecommunity-1-037-lb\">CollaboratingWithKaggleCommunity (1.037 LB)</a> predictions there are few outliers in the same place:\n<img src=\"https://i.imgur.com/ScY98U4.png\" alt=\"enter image description here\">\nGraphs for mjd differences and object ids:\n<img src=\"https://i.imgur.com/kabFlbv.png\" alt=\"enter image description here\">\n<img src=\"https://i.imgur.com/tSna7fm.png\" alt=\"enter image description here\">\nCould it be some kind of flux leakage or it's just my mistake?</p>",
      "rawMarkdown": "I used Scirpus method to predict class 99 and sorted predictions and metadata by object_id. Then I plotted mjd differences for objects where prediction for class 99 is maximum and their indices and found something strange. \nFor cnn+gru model:\n![enter image description here][1]\n\n\nFor [CollaboratingWithKaggleCommunity (1.037 LB)][2] predictions there are few outliers in the same place:\n![enter image description here][3]\nGraphs for mjd differences and object ids:\n![enter image description here][4]\n![enter image description here][5]\nCould it be some kind of flux leakage or it's just my mistake?\n\n\n  [1]: https://i.imgur.com/FxRd4ue.png\n  [2]: https://www.kaggle.com/jimpsull/collaboratingwithkagglecommunity-1-037-lb\n  [3]: https://i.imgur.com/ScY98U4.png\n  [4]: https://i.imgur.com/kabFlbv.png\n  [5]: https://i.imgur.com/tSna7fm.png",
      "votes": null
    },
    {
      "id": "441577",
      "postDate": "12/18/2018 19:55:34",
      "content": "<p>I don't see any evidence of this in my predictions, the class 99 predictions (and every other one) have no difference near those object ids. The one noticeable difference is that the objects in the DDF have different properties from those outside of it. By default, the DDF objects are the first object ids in the test set. Did you somehow shift those to the middle?</p>",
      "rawMarkdown": "I don't see any evidence of this in my predictions, the class 99 predictions (and every other one) have no difference near those object ids. The one noticeable difference is that the objects in the DDF have different properties from those outside of it. By default, the DDF objects are the first object ids in the test set. Did you somehow shift those to the middle?",
      "votes": null
    },
    {
      "id": "441645",
      "postDate": "12/18/2018 22:09:40",
      "content": "<p>It seems DDF is not a problem:\n<img src=\"https://i.imgur.com/uHAKOln.png\" alt=\"Image\">\nThe column appears when my model predicts a high value for galactic class when it's exagalactic, so when I make them equal 0, class 99 become maximum\n<img src=\"https://i.imgur.com/vFwtWZo.png\" alt=\"Image\">\nbut for public predictions, it's something else\n<img src=\"https://i.imgur.com/9X8GyyN.png\" alt=\"Image\"></p>",
      "rawMarkdown": "It seems DDF is not a problem:\n![Image](https://i.imgur.com/uHAKOln.png)\nThe column appears when my model predicts a high value for galactic class when it's exagalactic, so when I make them equal 0, class 99 become maximum\n![Image](https://i.imgur.com/vFwtWZo.png)\nbut for public predictions, it's something else\n![Image](https://i.imgur.com/9X8GyyN.png)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 441577,
      "author_name": "kyleboone",
      "author_url": "",
      "post_date": "12/18/2018 19:55:34",
      "content": "<p>I don't see any evidence of this in my predictions, the class 99 predictions (and every other one) have no difference near those object ids. The one noticeable difference is that the objects in the DDF have different properties from those outside of it. By default, the DDF objects are the first object ids in the test set. Did you somehow shift those to the middle?</p>",
      "votes": null,
      "replies": [
        {
          "id": 441645,
          "author_name": "lrunaways",
          "author_url": "",
          "post_date": "12/18/2018 22:09:40",
          "content": "<p>It seems DDF is not a problem:\n<img src=\"https://i.imgur.com/uHAKOln.png\" alt=\"Image\">\nThe column appears when my model predicts a high value for galactic class when it's exagalactic, so when I make them equal 0, class 99 become maximum\n<img src=\"https://i.imgur.com/vFwtWZo.png\" alt=\"Image\">\nbut for public predictions, it's something else\n<img src=\"https://i.imgur.com/9X8GyyN.png\" alt=\"Image\"></p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "440974": "I used Scirpus method to predict class 99 and sorted predictions and metadata by object_id. Then I plotted mjd differences for objects where prediction for class 99 is maximum and their indices and found something strange. \nFor cnn+gru model:\n![enter image description here][1]\n\n\nFor [CollaboratingWithKaggleCommunity (1.037 LB)][2] predictions there are few outliers in the same place:\n![enter image description here][3]\nGraphs for mjd differences and object ids:\n![enter image description here][4]\n![enter image description here][5]\nCould it be some kind of flux leakage or it's just my mistake?\n\n\n  [1]: https://i.imgur.com/FxRd4ue.png\n  [2]: https://www.kaggle.com/jimpsull/collaboratingwithkagglecommunity-1-037-lb\n  [3]: https://i.imgur.com/ScY98U4.png\n  [4]: https://i.imgur.com/kabFlbv.png\n  [5]: https://i.imgur.com/tSna7fm.png",
    "441577": "I don't see any evidence of this in my predictions, the class 99 predictions (and every other one) have no difference near those object ids. The one noticeable difference is that the objects in the DDF have different properties from those outside of it. By default, the DDF objects are the first object ids in the test set. Did you somehow shift those to the middle?",
    "441645": "It seems DDF is not a problem:\n![Image](https://i.imgur.com/uHAKOln.png)\nThe column appears when my model predicts a high value for galactic class when it's exagalactic, so when I make them equal 0, class 99 become maximum\n![Image](https://i.imgur.com/vFwtWZo.png)\nbut for public predictions, it's something else\n![Image](https://i.imgur.com/9X8GyyN.png)"
  },
  "source": "meta"
}