{
  "id": 122270,
  "title": "Adversarial ASHRAE Redoux - It's Huge! [AUC = 0.96]",
  "url": "/competitions/ashrae-energy-prediction/discussion/122270",
  "author_name": "Bojan Tunguz",
  "post_date": "2019-12-19T00:50:30.762000",
  "votes": 30,
  "comment_count": 15,
  "views": 0,
  "content": "<p>Early on in the competition I tried to create an <a href=\"https://www.kaggle.com/tunguz/adversarial-ashrae\">adversarial notebook</a>.  However, due to the size of the dataset I was limited to just a few features and very basic LightGBM model. I wanted to make a better model offline, and share the results here. Unfortunately, I've been really busy lately with some major changes in my professional life, and only today I got a bit of time to do this small project. Hopefully it's not too late, and it might be useful to some of you. </p>\n\n<p>You can find the notebook on GitHub: <a href=\"https://github.com/tunguz/ASHRAE/blob/master/Adversarial.ipynb\">https://github.com/tunguz/ASHRAE/blob/master/Adversarial.ipynb</a></p>\n\n<p><img src=\"https://raw.githubusercontent.com/tunguz/ASHRAE/master/ASHRAE_feat_imp.png\" alt=\"feat imp\"></p>",
  "messages": [
    {
      "id": 698226,
      "postDate": "2019-12-19T00:50:30.763Z",
      "content": "<p>Early on in the competition I tried to create an <a href=\"https://www.kaggle.com/tunguz/adversarial-ashrae\">adversarial notebook</a>.  However, due to the size of the dataset I was limited to just a few features and very basic LightGBM model. I wanted to make a better model offline, and share the results here. Unfortunately, I've been really busy lately with some major changes in my professional life, and only today I got a bit of time to do this small project. Hopefully it's not too late, and it might be useful to some of you. </p>\n\n<p>You can find the notebook on GitHub: <a href=\"https://github.com/tunguz/ASHRAE/blob/master/Adversarial.ipynb\">https://github.com/tunguz/ASHRAE/blob/master/Adversarial.ipynb</a></p>\n\n<p><img src=\"https://raw.githubusercontent.com/tunguz/ASHRAE/master/ASHRAE_feat_imp.png\" alt=\"feat imp\"></p>",
      "rawMarkdown": "Early on in the competition I tried to create an [adversarial notebook](https://www.kaggle.com/tunguz/adversarial-ashrae).  However, due to the size of the dataset I was limited to just a few features and very basic LightGBM model. I wanted to make a better model offline, and share the results here. Unfortunately, I've been really busy lately with some major changes in my professional life, and only today I got a bit of time to do this small project. Hopefully it's not too late, and it might be useful to some of you. \n\nYou can find the notebook on GitHub: https://github.com/tunguz/ASHRAE/blob/master/Adversarial.ipynb\n\n![feat imp](https://raw.githubusercontent.com/tunguz/ASHRAE/master/ASHRAE_feat_imp.png)",
      "votes": 30
    },
    {
      "id": 698378,
      "postDate": "2019-12-19T06:34:07.883Z",
      "content": "<p>Thanks for trying it out. According to LOFO, wind speed doesn't have any negative or positive impact. My model without it also performed similar on LB. It may be related to how the feature is preprocessed, my adversarial scores are also not even close to 0.96.</p>",
      "rawMarkdown": "Thanks for trying it out. According to LOFO, wind speed doesn't have any negative or positive impact. My model without it also performed similar on LB. It may be related to how the feature is preprocessed, my adversarial scores are also not even close to 0.96.",
      "votes": 4
    },
    {
      "id": 698884,
      "postDate": "2019-12-19T20:27:00.887Z",
      "content": "<p>this could be... a hurricane!</p>",
      "rawMarkdown": "this could be... a hurricane!",
      "votes": 1
    },
    {
      "id": 698807,
      "postDate": "2019-12-19T17:56:14.737Z",
      "content": "<p>This is also consistent with what I found. I had it dropped although it looks counter-intuitive when wind_direction is still somewhat helpful. Thanks for sharing.</p>",
      "rawMarkdown": "This is also consistent with what I found. I had it dropped although it looks counter-intuitive when wind_direction is still somewhat helpful. Thanks for sharing.",
      "votes": 1
    },
    {
      "id": 698339,
      "postDate": "2019-12-19T05:08:49.260Z",
      "content": "<p>Good stuff, thanks for sharing!</p>\n\n<p>Congrats on the new job <a href=\"/tunguz\">@tunguz</a> </p>",
      "rawMarkdown": "Good stuff, thanks for sharing!\n\nCongrats on the new job @tunguz ",
      "votes": 1
    },
    {
      "id": 698296,
      "postDate": "2019-12-19T03:15:48.913Z",
      "content": "<p>congrats on the new job and thank you for the tip.</p>",
      "rawMarkdown": "congrats on the new job and thank you for the tip.",
      "votes": 1
    },
    {
      "id": 698292,
      "postDate": "2019-12-19T02:57:05.533Z",
      "content": "<p>Thanks for confirming that dropping wind_speed was for the best</p>",
      "rawMarkdown": "Thanks for confirming that dropping wind_speed was for the best",
      "votes": 1
    },
    {
      "id": 698485,
      "postDate": "2019-12-19T10:01:57.477Z",
      "content": "<p>Oh, wow, that's indeed a huge finding! Thank you <a href=\"/tunguz\">@tunguz</a> \nLB shakeup is unavoidable :(</p>",
      "rawMarkdown": "Oh, wow, that's indeed a huge finding! Thank you @tunguz \nLB shakeup is unavoidable :(",
      "votes": 2
    },
    {
      "id": 698228,
      "postDate": "2019-12-19T01:06:06.660Z",
      "content": "<p>Wow, thats insane!</p>",
      "rawMarkdown": "Wow, thats insane!",
      "votes": 2
    },
    {
      "id": 706046,
      "postDate": "2019-12-29T20:54:54.940Z",
      "content": "<p><a href=\"/tunguz\">@tunguz</a>  Thanks for sharing. Congratulations on your top 10 ranking under 3 categories</p>",
      "rawMarkdown": "@tunguz  Thanks for sharing. Congratulations on your top 10 ranking under 3 categories"
    },
    {
      "id": 701372,
      "postDate": "2019-12-23T12:14:57.160Z",
      "content": "<p>Good stuff</p>",
      "rawMarkdown": "Good stuff"
    },
    {
      "id": 698968,
      "postDate": "2019-12-20T00:17:14.720Z",
      "content": "<p>I tried but it didn't improve my score</p>",
      "rawMarkdown": "I tried but it didn't improve my score"
    },
    {
      "id": 700773,
      "postDate": "2019-12-22T15:22:16.383Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 698330,
      "postDate": "2019-12-19T04:44:06.240Z",
      "rawMarkdown": "",
      "votes": -2,
      "isDeleted": true
    },
    {
      "id": 698526,
      "postDate": "2019-12-19T10:55:58.937Z",
      "content": "<p>thanks，nice job！</p>",
      "rawMarkdown": "thanks，nice job！",
      "votes": 1
    },
    {
      "id": 698274,
      "postDate": "2019-12-19T02:18:55.233Z",
      "content": "<p>Good stuff, thanks for sharing!</p>",
      "rawMarkdown": "Good stuff, thanks for sharing!",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 698378,
      "author_name": "Ahmet Erdem",
      "author_url": "",
      "post_date": "2019-12-19T06:34:07.883000",
      "content": "<p>Thanks for trying it out. According to LOFO, wind speed doesn't have any negative or positive impact. My model without it also performed similar on LB. It may be related to how the feature is preprocessed, my adversarial scores are also not even close to 0.96.</p>",
      "votes": 4,
      "replies": []
    },
    {
      "id": 698884,
      "author_name": "Mycroft Holmes",
      "author_url": "",
      "post_date": "2019-12-19T20:27:00.887000",
      "content": "<p>this could be... a hurricane!</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 698807,
      "author_name": "Kevin Liu",
      "author_url": "",
      "post_date": "2019-12-19T17:56:14.737000",
      "content": "<p>This is also consistent with what I found. I had it dropped although it looks counter-intuitive when wind_direction is still somewhat helpful. Thanks for sharing.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 698339,
      "author_name": "Ailurophile",
      "author_url": "",
      "post_date": "2019-12-19T05:08:49.260000",
      "content": "<p>Good stuff, thanks for sharing!</p>\n\n<p>Congrats on the new job <a href=\"/tunguz\">@tunguz</a> </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 698296,
      "author_name": "eagle4",
      "author_url": "",
      "post_date": "2019-12-19T03:15:48.913000",
      "content": "<p>congrats on the new job and thank you for the tip.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 698292,
      "author_name": "yukiya",
      "author_url": "",
      "post_date": "2019-12-19T02:57:05.533000",
      "content": "<p>Thanks for confirming that dropping wind_speed was for the best</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 698485,
      "author_name": "illia kavalevich",
      "author_url": "",
      "post_date": "2019-12-19T10:01:57.477000",
      "content": "<p>Oh, wow, that's indeed a huge finding! Thank you <a href=\"/tunguz\">@tunguz</a> \nLB shakeup is unavoidable :(</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 698228,
      "author_name": "Carlo",
      "author_url": "",
      "post_date": "2019-12-19T01:06:06.660000",
      "content": "<p>Wow, thats insane!</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 706046,
      "author_name": "Tensor Girl",
      "author_url": "",
      "post_date": "2019-12-29T20:54:54.940000",
      "content": "<p><a href=\"/tunguz\">@tunguz</a>  Thanks for sharing. Congratulations on your top 10 ranking under 3 categories</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 701372,
      "author_name": "wei lu",
      "author_url": "",
      "post_date": "2019-12-23T12:14:57.160000",
      "content": "<p>Good stuff</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 698968,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-12-20T00:17:14.720000",
      "content": "<p>I tried but it didn't improve my score</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 700773,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-12-22T15:22:16.383000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 698330,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-12-19T04:44:06.240000",
      "content": "",
      "votes": -2,
      "replies": []
    },
    {
      "id": 698526,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-12-19T10:55:58.937000",
      "content": "<p>thanks，nice job！</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 698274,
      "author_name": "TJ Klein",
      "author_url": "",
      "post_date": "2019-12-19T02:18:55.233000",
      "content": "<p>Good stuff, thanks for sharing!</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "698226": "Early on in the competition I tried to create an [adversarial notebook](https://www.kaggle.com/tunguz/adversarial-ashrae).  However, due to the size of the dataset I was limited to just a few features and very basic LightGBM model. I wanted to make a better model offline, and share the results here. Unfortunately, I've been really busy lately with some major changes in my professional life, and only today I got a bit of time to do this small project. Hopefully it's not too late, and it might be useful to some of you. \n\nYou can find the notebook on GitHub: https://github.com/tunguz/ASHRAE/blob/master/Adversarial.ipynb\n\n![feat imp](https://raw.githubusercontent.com/tunguz/ASHRAE/master/ASHRAE_feat_imp.png)",
    "698378": "Thanks for trying it out. According to LOFO, wind speed doesn't have any negative or positive impact. My model without it also performed similar on LB. It may be related to how the feature is preprocessed, my adversarial scores are also not even close to 0.96.",
    "698884": "this could be... a hurricane!",
    "698807": "This is also consistent with what I found. I had it dropped although it looks counter-intuitive when wind_direction is still somewhat helpful. Thanks for sharing.",
    "698339": "Good stuff, thanks for sharing!\n\nCongrats on the new job @tunguz ",
    "698296": "congrats on the new job and thank you for the tip.",
    "698292": "Thanks for confirming that dropping wind_speed was for the best",
    "698485": "Oh, wow, that's indeed a huge finding! Thank you @tunguz \nLB shakeup is unavoidable :(",
    "698228": "Wow, thats insane!",
    "706046": "@tunguz  Thanks for sharing. Congratulations on your top 10 ranking under 3 categories",
    "701372": "Good stuff",
    "698968": "I tried but it didn't improve my score",
    "700773": "",
    "698330": "",
    "698526": "thanks，nice job！",
    "698274": "Good stuff, thanks for sharing!"
  }
}