{
  "id": 73474,
  "title": "Teaming up",
  "url": "/competitions/PLAsTiCC-2018/discussion/73474",
  "author_name": "",
  "post_date": "2018-12-03T13:27:01.920111100Z",
  "votes": 7,
  "comment_count": 10,
  "views": 0,
  "content": "<p>Hello everyone,</p>\n\n<p>The end is approaching and I'm struggling to keep motivated. I'm looking for a person/team to exchange ideas on what works and what doesn't. My best model is two separate LightGBMs with a around 100 features. I have also managed to calculate the Hessian of the competition loss function analytically, which gave me a 0.08 boost.</p>\n\n<p>Feel free to reach on Kaggle or by email: maxhalford25@gmail.com.</p>\n\n<p>EDIT: I prefer to only team up with people who have a score above the best kernel score, it's proof you've done some thinking of yourself.</p>\n\n<p>EDIT: I've teamed up with <a href=\"/iprapas\">@iprapas</a>. We'll first see how it goes between the two of us and then we'll consider bringing in more people.</p>",
  "messages": [
    {
      "id": "432155",
      "postDate": "12/03/2018 13:27:01",
      "content": "<p>Hello everyone,</p>\n\n<p>The end is approaching and I'm struggling to keep motivated. I'm looking for a person/team to exchange ideas on what works and what doesn't. My best model is two separate LightGBMs with a around 100 features. I have also managed to calculate the Hessian of the competition loss function analytically, which gave me a 0.08 boost.</p>\n\n<p>Feel free to reach on Kaggle or by email: maxhalford25@gmail.com.</p>\n\n<p>EDIT: I prefer to only team up with people who have a score above the best kernel score, it's proof you've done some thinking of yourself.</p>\n\n<p>EDIT: I've teamed up with <a href=\"/iprapas\">@iprapas</a>. We'll first see how it goes between the two of us and then we'll consider bringing in more people.</p>",
      "rawMarkdown": "Hello everyone,\n\nThe end is approaching and I'm struggling to keep motivated. I'm looking for a person/team to exchange ideas on what works and what doesn't. My best model is two separate LightGBMs with a around 100 features. I have also managed to calculate the Hessian of the competition loss function analytically, which gave me a 0.08 boost.\n\nFeel free to reach on Kaggle or by email: maxhalford25@gmail.com.\n\nEDIT: I prefer to only team up with people who have a score above the best kernel score, it's proof you've done some thinking of yourself.\n\nEDIT: I've teamed up with @iprapas. We'll first see how it goes between the two of us and then we'll consider bringing in more people.",
      "votes": null
    },
    {
      "id": "432175",
      "postDate": "12/03/2018 14:11:57",
      "content": "<p>I am interested. This is the confusion matrix of my best single LGBM model: <a href=\"https://www.kaggle.com/c/PLAsTiCC-2018/discussion/72613#429895\">https://www.kaggle.com/c/PLAsTiCC-2018/discussion/72613#429895</a></p>",
      "rawMarkdown": "I am interested. This is the confusion matrix of my best single LGBM model: https://www.kaggle.com/c/PLAsTiCC-2018/discussion/72613#429895",
      "votes": null
    },
    {
      "id": "432203",
      "postDate": "12/03/2018 14:51:56",
      "content": "<p>I am interested. I am studying astronomy but not supernova field. My LB score is 1.080 now. </p>",
      "rawMarkdown": "I am interested. I am studying astronomy but not supernova field. My LB score is 1.080 now.",
      "votes": null
    },
    {
      "id": "432235",
      "postDate": "12/03/2018 15:34:59",
      "content": "<p>I sent you a mail.</p>",
      "rawMarkdown": "I sent you a mail.",
      "votes": null
    },
    {
      "id": "432236",
      "postDate": "12/03/2018 15:35:51",
      "content": "<p>Hey, sorry but I prefer it if your best score is higher than the best kernel score, it shows you did some thinking of yourself. </p>",
      "rawMarkdown": "Hey, sorry but I prefer it if your best score is higher than the best kernel score, it shows you did some thinking of yourself.",
      "votes": null
    },
    {
      "id": "432277",
      "postDate": "12/03/2018 16:52:01",
      "content": "<p>I am also interested. Current score is the result of single LGBM</p>",
      "rawMarkdown": "I am also interested. Current score is the result of single LGBM",
      "votes": null
    },
    {
      "id": "432293",
      "postDate": "12/03/2018 17:25:13",
      "content": "<p>I'll send you an email :)</p>",
      "rawMarkdown": "I'll send you an email :)",
      "votes": null
    },
    {
      "id": "432355",
      "postDate": "12/03/2018 18:53:54",
      "content": "<p>Hi Max, if its not too much can you please elaborate little more on how you have used Hessian of loss function ?</p>",
      "rawMarkdown": "Hi Max, if its not too much can you please elaborate little more on how you have used Hessian of loss function ?",
      "votes": null
    },
    {
      "id": "432727",
      "postDate": "12/04/2018 08:58:56",
      "content": "<p>Hi Max, I'am also at lost for quite some time now and finding hard time discovering new ideas. Maybe we can share ideas</p>",
      "rawMarkdown": "Hi Max, I'am also at lost for quite some time now and finding hard time discovering new ideas. Maybe we can share ideas",
      "votes": null
    },
    {
      "id": "433658",
      "postDate": "12/05/2018 09:57:46",
      "content": "<p>Hello everyone,</p>\n\n<p>I'm getting a lot of emails and I'm pretty busy at work at the moment, I apologize if I didn't get back to you. iprapas and I are having a good time. We're looking for a person/team with a fairly good neural network model to add some diversity to our solution.</p>\n\n<p>Thank you to everyone for your interest. It's motivating!</p>",
      "rawMarkdown": "Hello everyone,\n\nI'm getting a lot of emails and I'm pretty busy at work at the moment, I apologize if I didn't get back to you. iprapas and I are having a good time. We're looking for a person/team with a fairly good neural network model to add some diversity to our solution.\n\nThank you to everyone for your interest. It's motivating!",
      "votes": null
    },
    {
      "id": "434443",
      "postDate": "12/06/2018 12:10:24",
      "content": "<p>I've used the Hessian of the loss function and given it to LightGBM, it turns out that this is somewhat equivalent to using appropriate sample weights. Specifically I've computed the Hessian thanks to the sympy library.</p>",
      "rawMarkdown": "I've used the Hessian of the loss function and given it to LightGBM, it turns out that this is somewhat equivalent to using appropriate sample weights. Specifically I've computed the Hessian thanks to the sympy library.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 432175,
      "author_name": "iprapas",
      "author_url": "",
      "post_date": "12/03/2018 14:11:57",
      "content": "<p>I am interested. This is the confusion matrix of my best single LGBM model: <a href=\"https://www.kaggle.com/c/PLAsTiCC-2018/discussion/72613#429895\">https://www.kaggle.com/c/PLAsTiCC-2018/discussion/72613#429895</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 432235,
          "author_name": "maxhalford",
          "author_url": "",
          "post_date": "12/03/2018 15:34:59",
          "content": "<p>I sent you a mail.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 432203,
      "author_name": "vinzzz",
      "author_url": "",
      "post_date": "12/03/2018 14:51:56",
      "content": "<p>I am interested. I am studying astronomy but not supernova field. My LB score is 1.080 now. </p>",
      "votes": null,
      "replies": [
        {
          "id": 432236,
          "author_name": "maxhalford",
          "author_url": "",
          "post_date": "12/03/2018 15:35:51",
          "content": "<p>Hey, sorry but I prefer it if your best score is higher than the best kernel score, it shows you did some thinking of yourself. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 432277,
      "author_name": "adityakumarsinha",
      "author_url": "",
      "post_date": "12/03/2018 16:52:01",
      "content": "<p>I am also interested. Current score is the result of single LGBM</p>",
      "votes": null,
      "replies": [
        {
          "id": 432293,
          "author_name": "maxhalford",
          "author_url": "",
          "post_date": "12/03/2018 17:25:13",
          "content": "<p>I'll send you an email :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 432355,
      "author_name": "mks2192",
      "author_url": "",
      "post_date": "12/03/2018 18:53:54",
      "content": "<p>Hi Max, if its not too much can you please elaborate little more on how you have used Hessian of loss function ?</p>",
      "votes": null,
      "replies": [
        {
          "id": 434443,
          "author_name": "maxhalford",
          "author_url": "",
          "post_date": "12/06/2018 12:10:24",
          "content": "<p>I've used the Hessian of the loss function and given it to LightGBM, it turns out that this is somewhat equivalent to using appropriate sample weights. Specifically I've computed the Hessian thanks to the sympy library.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 432727,
      "author_name": "niclasdoce",
      "author_url": "",
      "post_date": "12/04/2018 08:58:56",
      "content": "<p>Hi Max, I'am also at lost for quite some time now and finding hard time discovering new ideas. Maybe we can share ideas</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 433658,
      "author_name": "maxhalford",
      "author_url": "",
      "post_date": "12/05/2018 09:57:46",
      "content": "<p>Hello everyone,</p>\n\n<p>I'm getting a lot of emails and I'm pretty busy at work at the moment, I apologize if I didn't get back to you. iprapas and I are having a good time. We're looking for a person/team with a fairly good neural network model to add some diversity to our solution.</p>\n\n<p>Thank you to everyone for your interest. It's motivating!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "432155": "Hello everyone,\n\nThe end is approaching and I'm struggling to keep motivated. I'm looking for a person/team to exchange ideas on what works and what doesn't. My best model is two separate LightGBMs with a around 100 features. I have also managed to calculate the Hessian of the competition loss function analytically, which gave me a 0.08 boost.\n\nFeel free to reach on Kaggle or by email: maxhalford25@gmail.com.\n\nEDIT: I prefer to only team up with people who have a score above the best kernel score, it's proof you've done some thinking of yourself.\n\nEDIT: I've teamed up with @iprapas. We'll first see how it goes between the two of us and then we'll consider bringing in more people.",
    "432175": "I am interested. This is the confusion matrix of my best single LGBM model: https://www.kaggle.com/c/PLAsTiCC-2018/discussion/72613#429895",
    "432203": "I am interested. I am studying astronomy but not supernova field. My LB score is 1.080 now.",
    "432235": "I sent you a mail.",
    "432236": "Hey, sorry but I prefer it if your best score is higher than the best kernel score, it shows you did some thinking of yourself.",
    "432277": "I am also interested. Current score is the result of single LGBM",
    "432293": "I'll send you an email :)",
    "432355": "Hi Max, if its not too much can you please elaborate little more on how you have used Hessian of loss function ?",
    "432727": "Hi Max, I'am also at lost for quite some time now and finding hard time discovering new ideas. Maybe we can share ideas",
    "433658": "Hello everyone,\n\nI'm getting a lot of emails and I'm pretty busy at work at the moment, I apologize if I didn't get back to you. iprapas and I are having a good time. We're looking for a person/team with a fairly good neural network model to add some diversity to our solution.\n\nThank you to everyone for your interest. It's motivating!",
    "434443": "I've used the Hessian of the loss function and given it to LightGBM, it turns out that this is somewhat equivalent to using appropriate sample weights. Specifically I've computed the Hessian thanks to the sympy library."
  },
  "source": "meta"
}