{
  "id": 25392,
  "title": "Some useful features",
  "url": "/competitions/bosch-production-line-performance/discussion/25392",
  "author_name": "",
  "post_date": "2016-11-13T12:30:16.223Z",
  "votes": 4,
  "comment_count": 2,
  "views": 434,
  "content": "<p>Here are some features I did not see mentioned in top solutions reports so far.  Take them as grain of salt as we didn't finish that well (partly due to late bug introduction from me), but here they are anyway;</p>\n\n<p>We assumed Beluga got it right, i.e. that for date features we have 0.1 = one hour.  We computed the time in day and time in week via modulo 2.4 and 16.8. more precisely:\ntime_day = (10 * time) % 24\ntime_week = (10 * time) % 168</p>\n\n<p>Then we computed out of fold binned response average on these.</p>\n\n<p>These were more poweful for us that faron's magic features.</p>",
  "messages": [
    {
      "id": "144259",
      "postDate": "11/13/2016 12:30:16",
      "content": "<p>Here are some features I did not see mentioned in top solutions reports so far.  Take them as grain of salt as we didn't finish that well (partly due to late bug introduction from me), but here they are anyway;</p>\n\n<p>We assumed Beluga got it right, i.e. that for date features we have 0.1 = one hour.  We computed the time in day and time in week via modulo 2.4 and 16.8. more precisely:\ntime_day = (10 * time) % 24\ntime_week = (10 * time) % 168</p>\n\n<p>Then we computed out of fold binned response average on these.</p>\n\n<p>These were more poweful for us that faron's magic features.</p>",
      "rawMarkdown": "Here are some features I did not see mentioned in top solutions reports so far.  Take them as grain of salt as we didn't finish that well (partly due to late bug introduction from me), but here they are anyway;\r\n\r\nWe assumed Beluga got it right, i.e. that for date features we have 0.1 = one hour.  We computed the time in day and time in week via modulo 2.4 and 16.8. more precisely:\r\ntime_day = (10 * time) % 24\r\ntime_week = (10 * time) % 168\r\n\r\nThen we computed out of fold binned response average on these.\r\n\r\nThese were more poweful for us that faron's magic features.",
      "votes": null
    },
    {
      "id": "144268",
      "postDate": "11/13/2016 13:45:40",
      "content": "<p>[quote=CPMP;144259]</p>\n\n<p>Here are some features I did not see mentioned in top solutions reports so far.  Take them as grain of salt as we didn't finish that well (partly due to late bug introduction from me), but here they are anyway;</p>\n\n<p>We assumed Beluga got it right, i.e. that for date features we have 0.1 = one hour.  We computed the time in day and time in week via modulo 2.4 and 16.8. more precisely:\ntime_day = (10 * time) % 24\ntime_week = (10 * time) % 168</p>\n\n<p>Then we computed out of fold binned response average on these.</p>\n\n<p>These were more poweful for us that faron's magic features.</p>\n\n<p>[/quote]</p>\n\n<p>Just wondering, when you meant more powerful, do you mean there were more splits for the variable, or there was higher gain. I assume you are using xgboost. I realized that faron's magic features provided the most gain by a very significant margin, though the number of splits(fscore) was top 5 or so.  </p>",
      "rawMarkdown": "[quote=CPMP;144259]\r\n\r\nHere are some features I did not see mentioned in top solutions reports so far.  Take them as grain of salt as we didn't finish that well (partly due to late bug introduction from me), but here they are anyway;\r\n\r\nWe assumed Beluga got it right, i.e. that for date features we have 0.1 = one hour.  We computed the time in day and time in week via modulo 2.4 and 16.8. more precisely:\r\ntime_day = (10 * time) % 24\r\ntime_week = (10 * time) % 168\r\n\r\nThen we computed out of fold binned response average on these.\r\n\r\nThese were more poweful for us that faron's magic features.\r\n\r\n\r\n[/quote]\r\n\r\nJust wondering, when you meant more powerful, do you mean there were more splits for the variable, or there was higher gain. I assume you are using xgboost. I realized that faron's magic features provided the most gain by a very significant margin, though the number of splits(fscore) was top 5 or so.",
      "votes": null
    },
    {
      "id": "144272",
      "postDate": "11/13/2016 14:16:52",
      "content": "<p>I mean feature importance according to xgboost.  In practice, I think faron's feature provided a larger boost on LB score than these features.</p>\n\n<p>We used the following features for the leak (used by all top solutions according to reports so far):</p>\n\n<ul>\n<li>id delta from previous and from next row when sorted by min time +id , or max time +id.  </li>\n<li>response from next and from previous rows with same orderings.</li>\n</ul>\n\n<p>Faron's original features ending in 3 and 4 correspond to the id delta when you use min time+ id ordering.</p>\n\n<p>I tried to use more orderings, eg per station time, but did not found them that useful.  Seems that they were useful to top teams, I have missed something here.</p>",
      "rawMarkdown": "I mean feature importance according to xgboost.  In practice, I think faron's feature provided a larger boost on LB score than these features.\r\n\r\nWe used the following features for the leak (used by all top solutions according to reports so far):\r\n\r\n- id delta from previous and from next row when sorted by min time +id , or max time +id.  \r\n- response from next and from previous rows with same orderings.\r\n\r\nFaron's original features ending in 3 and 4 correspond to the id delta when you use min time+ id ordering.\r\n\r\nI tried to use more orderings, eg per station time, but did not found them that useful.  Seems that they were useful to top teams, I have missed something here.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 144268,
      "author_name": "leexiongan",
      "author_url": "",
      "post_date": "11/13/2016 13:45:40",
      "content": "<p>[quote=CPMP;144259]</p>\n\n<p>Here are some features I did not see mentioned in top solutions reports so far.  Take them as grain of salt as we didn't finish that well (partly due to late bug introduction from me), but here they are anyway;</p>\n\n<p>We assumed Beluga got it right, i.e. that for date features we have 0.1 = one hour.  We computed the time in day and time in week via modulo 2.4 and 16.8. more precisely:\ntime_day = (10 * time) % 24\ntime_week = (10 * time) % 168</p>\n\n<p>Then we computed out of fold binned response average on these.</p>\n\n<p>These were more poweful for us that faron's magic features.</p>\n\n<p>[/quote]</p>\n\n<p>Just wondering, when you meant more powerful, do you mean there were more splits for the variable, or there was higher gain. I assume you are using xgboost. I realized that faron's magic features provided the most gain by a very significant margin, though the number of splits(fscore) was top 5 or so.  </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 144272,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "11/13/2016 14:16:52",
      "content": "<p>I mean feature importance according to xgboost.  In practice, I think faron's feature provided a larger boost on LB score than these features.</p>\n\n<p>We used the following features for the leak (used by all top solutions according to reports so far):</p>\n\n<ul>\n<li>id delta from previous and from next row when sorted by min time +id , or max time +id.  </li>\n<li>response from next and from previous rows with same orderings.</li>\n</ul>\n\n<p>Faron's original features ending in 3 and 4 correspond to the id delta when you use min time+ id ordering.</p>\n\n<p>I tried to use more orderings, eg per station time, but did not found them that useful.  Seems that they were useful to top teams, I have missed something here.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "144259": "Here are some features I did not see mentioned in top solutions reports so far.  Take them as grain of salt as we didn't finish that well (partly due to late bug introduction from me), but here they are anyway;\r\n\r\nWe assumed Beluga got it right, i.e. that for date features we have 0.1 = one hour.  We computed the time in day and time in week via modulo 2.4 and 16.8. more precisely:\r\ntime_day = (10 * time) % 24\r\ntime_week = (10 * time) % 168\r\n\r\nThen we computed out of fold binned response average on these.\r\n\r\nThese were more poweful for us that faron's magic features.",
    "144268": "[quote=CPMP;144259]\r\n\r\nHere are some features I did not see mentioned in top solutions reports so far.  Take them as grain of salt as we didn't finish that well (partly due to late bug introduction from me), but here they are anyway;\r\n\r\nWe assumed Beluga got it right, i.e. that for date features we have 0.1 = one hour.  We computed the time in day and time in week via modulo 2.4 and 16.8. more precisely:\r\ntime_day = (10 * time) % 24\r\ntime_week = (10 * time) % 168\r\n\r\nThen we computed out of fold binned response average on these.\r\n\r\nThese were more poweful for us that faron's magic features.\r\n\r\n\r\n[/quote]\r\n\r\nJust wondering, when you meant more powerful, do you mean there were more splits for the variable, or there was higher gain. I assume you are using xgboost. I realized that faron's magic features provided the most gain by a very significant margin, though the number of splits(fscore) was top 5 or so.",
    "144272": "I mean feature importance according to xgboost.  In practice, I think faron's feature provided a larger boost on LB score than these features.\r\n\r\nWe used the following features for the leak (used by all top solutions according to reports so far):\r\n\r\n- id delta from previous and from next row when sorted by min time +id , or max time +id.  \r\n- response from next and from previous rows with same orderings.\r\n\r\nFaron's original features ending in 3 and 4 correspond to the id delta when you use min time+ id ordering.\r\n\r\nI tried to use more orderings, eg per station time, but did not found them that useful.  Seems that they were useful to top teams, I have missed something here."
  },
  "source": "meta"
}