{
  "id": 74829,
  "title": "Write how you solved the problem. Even if you're in the 200+ place [AFTER DEADLINE]",
  "url": "/competitions/PLAsTiCC-2018/discussion/74829",
  "author_name": "",
  "post_date": "2018-12-16T09:17:46.782685900Z",
  "votes": 8,
  "comment_count": 18,
  "views": 0,
  "content": "<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/439755/10883/gusi_lebedi760.jpg\" alt=\"Trying not to talk about your experience to young geese\">\nPlease, Write how you solved the problem. Even if you're in the 200+ place (<strong>after deadline</strong>)</p>\n\n<p>Traditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution.</p>\n\n<p>However, for my opinion this is absolutely not enough for the development of the community, best practices and understanding the whole picture of what is happening in the competition.</p>\n\n<p>The main problem is that from the best write-ups we learn what needs to be done to take a high position. However, behind the scenes remains:</p>\n\n<ul>\n<li>What did not have to do to be in the top</li>\n<li>Where on LB are those who have done less and significantly less than those who were in the top. It is possible that among these solutions there are really good.</li>\n</ul>\n\n<p>Therefore, if you are reading this post - describe your decision. Start with description of your background. Write WHAT you did. And not least, WHY you did just that. Describe the way you came to your solution. If on this way there were suddely insights, then describe them. Do a little analysis of why you are in the current location, and not higher. Such an analysis is primarily useful to you.</p>\n\n<p>If this is possible, share code your solution.</p>\n\n<p>If you worked in a team, describe how you communicated. Write a couple of sentences about how you manage code project.</p>\n\n<p>If you speak English poorly - use an online translator - We will understand you.</p>\n\n<p>Thank you and Good Luck</p>",
  "messages": [
    {
      "id": "439755",
      "postDate": "12/16/2018 09:17:46",
      "content": "<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/439755/10883/gusi_lebedi760.jpg\" alt=\"Trying not to talk about your experience to young geese\">\nPlease, Write how you solved the problem. Even if you're in the 200+ place (<strong>after deadline</strong>)</p>\n\n<p>Traditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution.</p>\n\n<p>However, for my opinion this is absolutely not enough for the development of the community, best practices and understanding the whole picture of what is happening in the competition.</p>\n\n<p>The main problem is that from the best write-ups we learn what needs to be done to take a high position. However, behind the scenes remains:</p>\n\n<ul>\n<li>What did not have to do to be in the top</li>\n<li>Where on LB are those who have done less and significantly less than those who were in the top. It is possible that among these solutions there are really good.</li>\n</ul>\n\n<p>Therefore, if you are reading this post - describe your decision. Start with description of your background. Write WHAT you did. And not least, WHY you did just that. Describe the way you came to your solution. If on this way there were suddely insights, then describe them. Do a little analysis of why you are in the current location, and not higher. Such an analysis is primarily useful to you.</p>\n\n<p>If this is possible, share code your solution.</p>\n\n<p>If you worked in a team, describe how you communicated. Write a couple of sentences about how you manage code project.</p>\n\n<p>If you speak English poorly - use an online translator - We will understand you.</p>\n\n<p>Thank you and Good Luck</p>",
      "rawMarkdown": "![Trying not to talk about your experience to young geese][1]\nPlease, Write how you solved the problem. Even if you're in the 200+ place (**after deadline**)\n\nTraditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution.\n\nHowever, for my opinion this is absolutely not enough for the development of the community, best practices and understanding the whole picture of what is happening in the competition.\n\nThe main problem is that from the best write-ups we learn what needs to be done to take a high position. However, behind the scenes remains:\n\n- What did not have to do to be in the top\n- Where on LB are those who have done less and significantly less than those who were in the top. It is possible that among these solutions there are really good.\n\nTherefore, if you are reading this post - describe your decision. Start with description of your background. Write WHAT you did. And not least, WHY you did just that. Describe the way you came to your solution. If on this way there were suddely insights, then describe them. Do a little analysis of why you are in the current location, and not higher. Such an analysis is primarily useful to you.\n\nIf this is possible, share code your solution.\n\nIf you worked in a team, describe how you communicated. Write a couple of sentences about how you manage code project.\n\nIf you speak English poorly - use an online translator - We will understand you.\n\nThank you and Good Luck\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/439755/10883/gusi_lebedi760.jpg",
      "votes": null
    },
    {
      "id": "439802",
      "postDate": "12/16/2018 11:56:55",
      "content": "<p>Traditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution. </p>\n\n<p>The reason for this is that this is the fairest way - it is extremely unfair to divulge ideas before the competition ends with only a couple of days to go  - so I down voted you! Sorry!</p>",
      "rawMarkdown": "Traditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution. \n\nThe reason for this is that this is the fairest way - it is extremely unfair to divulge ideas before the competition ends with only a couple of days to go  - so I down voted you! Sorry!",
      "votes": null
    },
    {
      "id": "439810",
      "postDate": "12/16/2018 12:17:22",
      "content": "<blockquote>\n  <p>Traditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution.</p>\n</blockquote>\n\n<p>I guess he wanted to emphasize \"the participants who demonstrate best result\" instead of \"after the end of the competition\". So I think he expects others to write their solutions but after the competition obviously. As a side note, I disagree with it. I think the best solutions and the solutions which are novel or time/memory/code length efficient are nice to share. If everyone shares the same things, we will have difficult time to find what is useful for us.</p>",
      "rawMarkdown": "&gt; Traditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution.\n\nI guess he wanted to emphasize \"the participants who demonstrate best result\" instead of \"after the end of the competition\". So I think he expects others to write their solutions but after the competition obviously. As a side note, I disagree with it. I think the best solutions and the solutions which are novel or time/memory/code length efficient are nice to share. If everyone shares the same things, we will have difficult time to find what is useful for us.",
      "votes": null
    },
    {
      "id": "439813",
      "postDate": "12/16/2018 12:33:16",
      "content": "<p>@Scirpus Nobody says to describe solution until the end of the competition:))</p>",
      "rawMarkdown": "Scirpus Nobody says to describe solution until the end of the competition:))",
      "votes": null
    },
    {
      "id": "439814",
      "postDate": "12/16/2018 12:37:16",
      "content": "<p>OK then I un-downvoted you!  So for everyone else answer Ruslan's stuff AFTER the competition ends!\nThanks for the clarification!</p>",
      "rawMarkdown": "OK then I un-downvoted you!  So for everyone else answer Ruslan's stuff AFTER the competition ends!\nThanks for the clarification!",
      "votes": null
    },
    {
      "id": "439820",
      "postDate": "12/16/2018 13:04:52",
      "content": "<p><a href=\"/divrikwicky\">@divrikwicky</a> \nThanks for the feedback. Let me suggest the following counterargument. We still expect the community to more value breakthrough write-up from the leaders. Such descriptions of course can be found on the tiles in the spirit of \"1/13/25 place solution ...\" and significantly more votes.</p>\n\n<p>&nbsp;In this sense, it seems that we will still be able to pick out good ideas from the general flow.</p>\n\n<p>My point is that the experience of participants who are not very high positions is also important. And unfortunately it is lost, because the participants believe that if they are not in the top, then no one is interested in finding out what methods they used to solve the problem.</p>\n\n<p>We all will be a little sad if really \"novel or time/memory/code length efficient \" will be part of just such an unpublished decision.</p>",
      "rawMarkdown": "divrikwicky \nThanks for the feedback. Let me suggest the following counterargument. We still expect the community to more value breakthrough write-up from the leaders. Such descriptions of course can be found on the tiles in the spirit of \"1/13/25 place solution ...\" and significantly more votes.\n\n&nbsp;In this sense, it seems that we will still be able to pick out good ideas from the general flow.\n\nMy point is that the experience of participants who are not very high positions is also important. And unfortunately it is lost, because the participants believe that if they are not in the top, then no one is interested in finding out what methods they used to solve the problem.\n\nWe all will be a little sad if really \"novel or time/memory/code length efficient \" will be part of just such an unpublished decision.",
      "votes": null
    },
    {
      "id": "439825",
      "postDate": "12/16/2018 13:28:00",
      "content": "<p>I agree, even when we finish in lower rankings, than say top 50, still there would be valuable insights  and experiences to share.</p>",
      "rawMarkdown": "I agree, even when we finish in lower rankings, than say top 50, still there would be valuable insights  and experiences to share.",
      "votes": null
    },
    {
      "id": "440060",
      "postDate": "12/17/2018 01:07:57",
      "content": "<p>Exactly.</p>",
      "rawMarkdown": "Exactly.",
      "votes": null
    },
    {
      "id": "440131",
      "postDate": "12/17/2018 05:30:16",
      "content": "<p>Interesting pic. </p>",
      "rawMarkdown": "Interesting pic.",
      "votes": null
    },
    {
      "id": "440173",
      "postDate": "12/17/2018 07:12:45",
      "content": "<p>I concur, writing down your experience helps you.  It helps you sort out what worked and didn't work, and it helps rememebring this for future use.  Helping others is agreat side effect, but it is not the main benefit for the one who writes IMHO.</p>\n\n<p>I know that the section \"what did not work\" will be way longer than the rest in our case ;)</p>",
      "rawMarkdown": "I concur, writing down your experience helps you.  It helps you sort out what worked and didn't work, and it helps rememebring this for future use.  Helping others is agreat side effect, but it is not the main benefit for the one who writes IMHO.\n\nI know that the section \"what did not work\" will be way longer than the rest in our case ;)",
      "votes": null
    },
    {
      "id": "440180",
      "postDate": "12/17/2018 07:16:46",
      "content": "<p>If possible, describe at least highlights of \"what did not work\" :)</p>",
      "rawMarkdown": "If possible, describe at least highlights of \"what did not work\" :)",
      "votes": null
    },
    {
      "id": "440182",
      "postDate": "12/17/2018 07:20:17",
      "content": "<p>Looking forward to see write-up on your experience. </p>",
      "rawMarkdown": "Looking forward to see write-up on your experience.",
      "votes": null
    },
    {
      "id": "440199",
      "postDate": "12/17/2018 07:55:32",
      "content": "<p>In this competition, there are many things we can do (unlike my previous competition, talkingdata). so I think it's possible 30th place solution includes something valuable the winner doesn't notice.</p>",
      "rawMarkdown": "In this competition, there are many things we can do (unlike my previous competition, talkingdata). so I think it's possible 30th place solution includes something valuable the winner doesn't notice.",
      "votes": null
    },
    {
      "id": "440220",
      "postDate": "12/17/2018 08:41:35",
      "content": "<blockquote>\n  <p>If possible, describe at least highlights of \"what did not work\" :)</p>\n</blockquote>\n\n<p>You're telling me?  I thought I said we will have a large section on it.</p>",
      "rawMarkdown": "&gt; If possible, describe at least highlights of \"what did not work\" :)\n\nYou're telling me?  I thought I said we will have a large section on it.",
      "votes": null
    },
    {
      "id": "440242",
      "postDate": "12/17/2018 09:26:16",
      "content": "<p><a href=\"/cpmpml\">@cpmpml</a> Have you already written your solution?</p>",
      "rawMarkdown": "cpmpml Have you already written your solution?",
      "votes": null
    },
    {
      "id": "440267",
      "postDate": "12/17/2018 10:00:14",
      "content": "<p>No, I'm still trying to improve it ;)  We haven't spent much time on blending, this is today's topic...  I doubt we'll catch up with you, but we will see.</p>",
      "rawMarkdown": "No, I'm still trying to improve it ;)  We haven't spent much time on blending, this is today's topic...  I doubt we'll catch up with you, but we will see.",
      "votes": null
    },
    {
      "id": "440739",
      "postDate": "12/18/2018 00:20:00",
      "content": "<p><a href=\"https://www.kaggle.com/jimpsull/plasticc-path-and-lessons-learned\">I put a brief summary of my journey here.</a>  The code is in the linked kernels.</p>\n\n<p>Cheers.</p>",
      "rawMarkdown": "[I put a brief summary of my journey here.](https://www.kaggle.com/jimpsull/plasticc-path-and-lessons-learned)  The code is in the linked kernels.\n\nCheers.",
      "votes": null
    },
    {
      "id": "441087",
      "postDate": "12/18/2018 08:46:13",
      "content": "<p>I just stole Olivier's Stuff for feature engineering - and did a mean of a NN, Genetic Program and LGB.</p>",
      "rawMarkdown": "I just stole Olivier's Stuff for feature engineering - and did a mean of a NN, Genetic Program and LGB.",
      "votes": null
    },
    {
      "id": "441330",
      "postDate": "12/18/2018 14:51:43",
      "content": "<p><strong>What didn't work (among lots of other things):</strong></p>\n\n<p>I ran a K-means algorithm with Dynamic Time Wrapping.  The underlying work I used for this is here:\n<a href=\"https://github.com/fpetitjean/DBA\">https://github.com/fpetitjean/DBA</a>\nVery cool stuff, maybe it will be useful for another project. It did make some nice curves and helped me see what classes were supernova or not. It also could differentiate between the easy classes; but not between e.g. 90 and 52, and not any better than the rest of my model.</p>\n\n<p><strong>What seemed to work:</strong></p>\n\n<p>SMOTE and DART helped with my overfitting issues</p>\n\n<p>Multiplying some features (e.g. max, amplitude) by hostgal_photoz^2</p>\n\n<p>Taking ratios of features between passbands </p>\n\n<p>Adjusting class weights to improve balance in confusion matrix (or at least that helped before using SMOTE/DART)</p>\n\n<p>Fitting a * np.exp(-b * x) on extragalactic objects, for each passband, for data after the object-level peak and until the last detected mjd (or maybe a bit later). I used the b (\"decay\") value and it seemed to help.</p>\n\n<p>** Impressions **</p>\n\n<p>This is maybe my 3rd or 4th competition, and the one I've spent the most energy on. It was also my favorite one. The data was interesting, and the project seemed to make some real-life sense and provide some value.</p>\n\n<p>The hard part was I felt I was flying blind much of the time. I first played around for a while without submitting. Then, towards the end, it would take me 1.5 hours to run dart LGBM, 1.5 hours to run predictions (even though my time consuming features were precomputed) , 45 minutes to submit (I reduced that to 12 seconds in the last few days). So, I was never really able to try a bunch of things and see how they affected the leaderboard. </p>\n\n<p>Thanks everyone, this was a lot of fun! Now it's time to go through all the posted solutions and try new ideas out!</p>",
      "rawMarkdown": "**What didn't work (among lots of other things):**\n\nI ran a K-means algorithm with Dynamic Time Wrapping.  The underlying work I used for this is here:\nhttps://github.com/fpetitjean/DBA\nVery cool stuff, maybe it will be useful for another project. It did make some nice curves and helped me see what classes were supernova or not. It also could differentiate between the easy classes; but not between e.g. 90 and 52, and not any better than the rest of my model.\n\n**What seemed to work:**\n\nSMOTE and DART helped with my overfitting issues\n\nMultiplying some features (e.g. max, amplitude) by hostgal_photoz^2\n\nTaking ratios of features between passbands \n\nAdjusting class weights to improve balance in confusion matrix (or at least that helped before using SMOTE/DART)\n\nFitting a * np.exp(-b * x) on extragalactic objects, for each passband, for data after the object-level peak and until the last detected mjd (or maybe a bit later). I used the b (\"decay\") value and it seemed to help.\n\n** Impressions **\n\nThis is maybe my 3rd or 4th competition, and the one I've spent the most energy on. It was also my favorite one. The data was interesting, and the project seemed to make some real-life sense and provide some value.\n\nThe hard part was I felt I was flying blind much of the time. I first played around for a while without submitting. Then, towards the end, it would take me 1.5 hours to run dart LGBM, 1.5 hours to run predictions (even though my time consuming features were precomputed) , 45 minutes to submit (I reduced that to 12 seconds in the last few days). So, I was never really able to try a bunch of things and see how they affected the leaderboard. \n\nThanks everyone, this was a lot of fun! Now it's time to go through all the posted solutions and try new ideas out!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 439802,
      "author_name": "scirpus",
      "author_url": "",
      "post_date": "12/16/2018 11:56:55",
      "content": "<p>Traditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution. </p>\n\n<p>The reason for this is that this is the fairest way - it is extremely unfair to divulge ideas before the competition ends with only a couple of days to go  - so I down voted you! Sorry!</p>",
      "votes": null,
      "replies": [
        {
          "id": 439810,
          "author_name": "aerdem4",
          "author_url": "",
          "post_date": "12/16/2018 12:17:22",
          "content": "<blockquote>\n  <p>Traditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution.</p>\n</blockquote>\n\n<p>I guess he wanted to emphasize \"the participants who demonstrate best result\" instead of \"after the end of the competition\". So I think he expects others to write their solutions but after the competition obviously. As a side note, I disagree with it. I think the best solutions and the solutions which are novel or time/memory/code length efficient are nice to share. If everyone shares the same things, we will have difficult time to find what is useful for us.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 439813,
          "author_name": "rooshroosh",
          "author_url": "",
          "post_date": "12/16/2018 12:33:16",
          "content": "<p>@Scirpus Nobody says to describe solution until the end of the competition:))</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 439814,
          "author_name": "scirpus",
          "author_url": "",
          "post_date": "12/16/2018 12:37:16",
          "content": "<p>OK then I un-downvoted you!  So for everyone else answer Ruslan's stuff AFTER the competition ends!\nThanks for the clarification!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 439820,
          "author_name": "rooshroosh",
          "author_url": "",
          "post_date": "12/16/2018 13:04:52",
          "content": "<p><a href=\"/divrikwicky\">@divrikwicky</a> \nThanks for the feedback. Let me suggest the following counterargument. We still expect the community to more value breakthrough write-up from the leaders. Such descriptions of course can be found on the tiles in the spirit of \"1/13/25 place solution ...\" and significantly more votes.</p>\n\n<p>&nbsp;In this sense, it seems that we will still be able to pick out good ideas from the general flow.</p>\n\n<p>My point is that the experience of participants who are not very high positions is also important. And unfortunately it is lost, because the participants believe that if they are not in the top, then no one is interested in finding out what methods they used to solve the problem.</p>\n\n<p>We all will be a little sad if really \"novel or time/memory/code length efficient \" will be part of just such an unpublished decision.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 439825,
          "author_name": "chinta",
          "author_url": "",
          "post_date": "12/16/2018 13:28:00",
          "content": "<p>I agree, even when we finish in lower rankings, than say top 50, still there would be valuable insights  and experiences to share.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440060,
          "author_name": "longyin2",
          "author_url": "",
          "post_date": "12/17/2018 01:07:57",
          "content": "<p>Exactly.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440199,
          "author_name": "mamasinkgs",
          "author_url": "",
          "post_date": "12/17/2018 07:55:32",
          "content": "<p>In this competition, there are many things we can do (unlike my previous competition, talkingdata). so I think it's possible 30th place solution includes something valuable the winner doesn't notice.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 440131,
      "author_name": "subrahmanyamv",
      "author_url": "",
      "post_date": "12/17/2018 05:30:16",
      "content": "<p>Interesting pic. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 440173,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "12/17/2018 07:12:45",
      "content": "<p>I concur, writing down your experience helps you.  It helps you sort out what worked and didn't work, and it helps rememebring this for future use.  Helping others is agreat side effect, but it is not the main benefit for the one who writes IMHO.</p>\n\n<p>I know that the section \"what did not work\" will be way longer than the rest in our case ;)</p>",
      "votes": null,
      "replies": [
        {
          "id": 440180,
          "author_name": "rooshroosh",
          "author_url": "",
          "post_date": "12/17/2018 07:16:46",
          "content": "<p>If possible, describe at least highlights of \"what did not work\" :)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440182,
          "author_name": "subrahmanyamv",
          "author_url": "",
          "post_date": "12/17/2018 07:20:17",
          "content": "<p>Looking forward to see write-up on your experience. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440220,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "12/17/2018 08:41:35",
          "content": "<blockquote>\n  <p>If possible, describe at least highlights of \"what did not work\" :)</p>\n</blockquote>\n\n<p>You're telling me?  I thought I said we will have a large section on it.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440242,
          "author_name": "mamasinkgs",
          "author_url": "",
          "post_date": "12/17/2018 09:26:16",
          "content": "<p><a href=\"/cpmpml\">@cpmpml</a> Have you already written your solution?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440267,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "12/17/2018 10:00:14",
          "content": "<p>No, I'm still trying to improve it ;)  We haven't spent much time on blending, this is today's topic...  I doubt we'll catch up with you, but we will see.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 440739,
      "author_name": "jimpsull",
      "author_url": "",
      "post_date": "12/18/2018 00:20:00",
      "content": "<p><a href=\"https://www.kaggle.com/jimpsull/plasticc-path-and-lessons-learned\">I put a brief summary of my journey here.</a>  The code is in the linked kernels.</p>\n\n<p>Cheers.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 441087,
      "author_name": "scirpus",
      "author_url": "",
      "post_date": "12/18/2018 08:46:13",
      "content": "<p>I just stole Olivier's Stuff for feature engineering - and did a mean of a NN, Genetic Program and LGB.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 441330,
      "author_name": "sdoria",
      "author_url": "",
      "post_date": "12/18/2018 14:51:43",
      "content": "<p><strong>What didn't work (among lots of other things):</strong></p>\n\n<p>I ran a K-means algorithm with Dynamic Time Wrapping.  The underlying work I used for this is here:\n<a href=\"https://github.com/fpetitjean/DBA\">https://github.com/fpetitjean/DBA</a>\nVery cool stuff, maybe it will be useful for another project. It did make some nice curves and helped me see what classes were supernova or not. It also could differentiate between the easy classes; but not between e.g. 90 and 52, and not any better than the rest of my model.</p>\n\n<p><strong>What seemed to work:</strong></p>\n\n<p>SMOTE and DART helped with my overfitting issues</p>\n\n<p>Multiplying some features (e.g. max, amplitude) by hostgal_photoz^2</p>\n\n<p>Taking ratios of features between passbands </p>\n\n<p>Adjusting class weights to improve balance in confusion matrix (or at least that helped before using SMOTE/DART)</p>\n\n<p>Fitting a * np.exp(-b * x) on extragalactic objects, for each passband, for data after the object-level peak and until the last detected mjd (or maybe a bit later). I used the b (\"decay\") value and it seemed to help.</p>\n\n<p>** Impressions **</p>\n\n<p>This is maybe my 3rd or 4th competition, and the one I've spent the most energy on. It was also my favorite one. The data was interesting, and the project seemed to make some real-life sense and provide some value.</p>\n\n<p>The hard part was I felt I was flying blind much of the time. I first played around for a while without submitting. Then, towards the end, it would take me 1.5 hours to run dart LGBM, 1.5 hours to run predictions (even though my time consuming features were precomputed) , 45 minutes to submit (I reduced that to 12 seconds in the last few days). So, I was never really able to try a bunch of things and see how they affected the leaderboard. </p>\n\n<p>Thanks everyone, this was a lot of fun! Now it's time to go through all the posted solutions and try new ideas out!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "439755": "![Trying not to talk about your experience to young geese][1]\nPlease, Write how you solved the problem. Even if you're in the 200+ place (**after deadline**)\n\nTraditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution.\n\nHowever, for my opinion this is absolutely not enough for the development of the community, best practices and understanding the whole picture of what is happening in the competition.\n\nThe main problem is that from the best write-ups we learn what needs to be done to take a high position. However, behind the scenes remains:\n\n- What did not have to do to be in the top\n- Where on LB are those who have done less and significantly less than those who were in the top. It is possible that among these solutions there are really good.\n\nTherefore, if you are reading this post - describe your decision. Start with description of your background. Write WHAT you did. And not least, WHY you did just that. Describe the way you came to your solution. If on this way there were suddely insights, then describe them. Do a little analysis of why you are in the current location, and not higher. Such an analysis is primarily useful to you.\n\nIf this is possible, share code your solution.\n\nIf you worked in a team, describe how you communicated. Write a couple of sentences about how you manage code project.\n\nIf you speak English poorly - use an online translator - We will understand you.\n\nThank you and Good Luck\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/439755/10883/gusi_lebedi760.jpg",
    "439802": "Traditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution. \n\nThe reason for this is that this is the fairest way - it is extremely unfair to divulge ideas before the competition ends with only a couple of days to go  - so I down voted you! Sorry!",
    "439810": "&gt; Traditionally, after the end of the competition, the participants who demonstrate best result share the description of their solution.\n\nI guess he wanted to emphasize \"the participants who demonstrate best result\" instead of \"after the end of the competition\". So I think he expects others to write their solutions but after the competition obviously. As a side note, I disagree with it. I think the best solutions and the solutions which are novel or time/memory/code length efficient are nice to share. If everyone shares the same things, we will have difficult time to find what is useful for us.",
    "439813": "Scirpus Nobody says to describe solution until the end of the competition:))",
    "439814": "OK then I un-downvoted you!  So for everyone else answer Ruslan's stuff AFTER the competition ends!\nThanks for the clarification!",
    "439820": "divrikwicky \nThanks for the feedback. Let me suggest the following counterargument. We still expect the community to more value breakthrough write-up from the leaders. Such descriptions of course can be found on the tiles in the spirit of \"1/13/25 place solution ...\" and significantly more votes.\n\n&nbsp;In this sense, it seems that we will still be able to pick out good ideas from the general flow.\n\nMy point is that the experience of participants who are not very high positions is also important. And unfortunately it is lost, because the participants believe that if they are not in the top, then no one is interested in finding out what methods they used to solve the problem.\n\nWe all will be a little sad if really \"novel or time/memory/code length efficient \" will be part of just such an unpublished decision.",
    "439825": "I agree, even when we finish in lower rankings, than say top 50, still there would be valuable insights  and experiences to share.",
    "440060": "Exactly.",
    "440131": "Interesting pic.",
    "440173": "I concur, writing down your experience helps you.  It helps you sort out what worked and didn't work, and it helps rememebring this for future use.  Helping others is agreat side effect, but it is not the main benefit for the one who writes IMHO.\n\nI know that the section \"what did not work\" will be way longer than the rest in our case ;)",
    "440180": "If possible, describe at least highlights of \"what did not work\" :)",
    "440182": "Looking forward to see write-up on your experience.",
    "440199": "In this competition, there are many things we can do (unlike my previous competition, talkingdata). so I think it's possible 30th place solution includes something valuable the winner doesn't notice.",
    "440220": "&gt; If possible, describe at least highlights of \"what did not work\" :)\n\nYou're telling me?  I thought I said we will have a large section on it.",
    "440242": "cpmpml Have you already written your solution?",
    "440267": "No, I'm still trying to improve it ;)  We haven't spent much time on blending, this is today's topic...  I doubt we'll catch up with you, but we will see.",
    "440739": "[I put a brief summary of my journey here.](https://www.kaggle.com/jimpsull/plasticc-path-and-lessons-learned)  The code is in the linked kernels.\n\nCheers.",
    "441087": "I just stole Olivier's Stuff for feature engineering - and did a mean of a NN, Genetic Program and LGB.",
    "441330": "**What didn't work (among lots of other things):**\n\nI ran a K-means algorithm with Dynamic Time Wrapping.  The underlying work I used for this is here:\nhttps://github.com/fpetitjean/DBA\nVery cool stuff, maybe it will be useful for another project. It did make some nice curves and helped me see what classes were supernova or not. It also could differentiate between the easy classes; but not between e.g. 90 and 52, and not any better than the rest of my model.\n\n**What seemed to work:**\n\nSMOTE and DART helped with my overfitting issues\n\nMultiplying some features (e.g. max, amplitude) by hostgal_photoz^2\n\nTaking ratios of features between passbands \n\nAdjusting class weights to improve balance in confusion matrix (or at least that helped before using SMOTE/DART)\n\nFitting a * np.exp(-b * x) on extragalactic objects, for each passband, for data after the object-level peak and until the last detected mjd (or maybe a bit later). I used the b (\"decay\") value and it seemed to help.\n\n** Impressions **\n\nThis is maybe my 3rd or 4th competition, and the one I've spent the most energy on. It was also my favorite one. The data was interesting, and the project seemed to make some real-life sense and provide some value.\n\nThe hard part was I felt I was flying blind much of the time. I first played around for a while without submitting. Then, towards the end, it would take me 1.5 hours to run dart LGBM, 1.5 hours to run predictions (even though my time consuming features were precomputed) , 45 minutes to submit (I reduced that to 12 seconds in the last few days). So, I was never really able to try a bunch of things and see how they affected the leaderboard. \n\nThanks everyone, this was a lot of fun! Now it's time to go through all the posted solutions and try new ideas out!"
  },
  "source": "meta"
}