{
  "id": 22045,
  "title": "Anyone predicted which place_id will diminish in the future?",
  "url": "/competitions/facebook-v-predicting-check-ins/discussion/22045",
  "author_name": "",
  "post_date": "2016-07-05T03:01:17.600Z",
  "votes": null,
  "comment_count": 4,
  "views": 449,
  "content": "<p>As title.</p>\n\n<p>Might be too late to ask. But since no one ask this question, I just want to ask if anyone did this, and whether it could improve LB score.</p>\n\n<p>If any similar work related to this, I remember someone claim a slight increase if randomly delete some of the most popular place_ids.</p>",
  "messages": [
    {
      "id": "125964",
      "postDate": "07/05/2016 03:01:17",
      "content": "<p>As title.</p>\n\n<p>Might be too late to ask. But since no one ask this question, I just want to ask if anyone did this, and whether it could improve LB score.</p>\n\n<p>If any similar work related to this, I remember someone claim a slight increase if randomly delete some of the most popular place_ids.</p>",
      "rawMarkdown": "As title.\r\n\r\nMight be too late to ask. But since no one ask this question, I just want to ask if anyone did this, and whether it could improve LB score.\r\n\r\nIf any similar work related to this, I remember someone claim a slight increase if randomly delete some of the most popular place_ids.",
      "votes": null
    },
    {
      "id": "125966",
      "postDate": "07/05/2016 03:26:20",
      "content": "<p>This idea is interesting. I am looking forward to hear any insight.</p>",
      "rawMarkdown": "This idea is interesting. I am looking forward to hear any insight.",
      "votes": null
    },
    {
      "id": "125967",
      "postDate": "07/05/2016 03:59:24",
      "content": "<p>@Li Li, Thanks for replying.</p>\n\n<p>I have tried to multiply a weight to the final probability generated by KNN before taking the top 3 ranks. The weight is just a rough estimate of whether the place id will <strong>increase</strong>, <strong>remain the same level</strong>, or <strong>decrease</strong>. The estimation is calculate by whether the number of check_ins increase or decrease, and the weight is <strong>1.1</strong>, <strong>1</strong>, and <strong>0.9</strong>.</p>\n\n<p>But sadly, the result did not improve. I guess the reason is because KNN already take the year parameter into account, and give recent year a closer distance.</p>\n\n<p>I think the idea of predicting diminish place_ids should intuitively work, but I am not sure how to better include this information into the final model.</p>\n\n<p>That's all my thoughts and findings.</p>",
      "rawMarkdown": "Li Li, Thanks for replying.\r\n\r\nI have tried to multiply a weight to the final probability generated by KNN before taking the top 3 ranks. The weight is just a rough estimate of whether the place id will **increase**, **remain the same level**, or **decrease**. The estimation is calculate by whether the number of check_ins increase or decrease, and the weight is **1.1**, **1**, and **0.9**.\r\n\r\nBut sadly, the result did not improve. I guess the reason is because KNN already take the year parameter into account, and give recent year a closer distance.\r\n\r\nI think the idea of predicting diminish place_ids should intuitively work, but I am not sure how to better include this information into the final model.\r\n\r\nThat's all my thoughts and findings.",
      "votes": null
    },
    {
      "id": "126016",
      "postDate": "07/05/2016 14:57:51",
      "content": "<p>I also tried this, using various forms of regression to identify places that were declining in popularity, but like you, my results did not improve.</p>",
      "rawMarkdown": "I also tried this, using various forms of regression to identify places that were declining in popularity, but like you, my results did not improve.",
      "votes": null
    },
    {
      "id": "126305",
      "postDate": "07/07/2016 15:52:01",
      "content": "<p>There are ~500 place_ids that showed up only during the 1st half of 1st year, while the test set is for the 2nd half of 2nd year. Deleting these got 0.00001 change in LB score. </p>",
      "rawMarkdown": "There are ~500 place_ids that showed up only during the 1st half of 1st year, while the test set is for the 2nd half of 2nd year. Deleting these got 0.00001 change in LB score.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 125966,
      "author_name": "aikinogard",
      "author_url": "",
      "post_date": "07/05/2016 03:26:20",
      "content": "<p>This idea is interesting. I am looking forward to hear any insight.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 125967,
      "author_name": "ma350365879",
      "author_url": "",
      "post_date": "07/05/2016 03:59:24",
      "content": "<p>@Li Li, Thanks for replying.</p>\n\n<p>I have tried to multiply a weight to the final probability generated by KNN before taking the top 3 ranks. The weight is just a rough estimate of whether the place id will <strong>increase</strong>, <strong>remain the same level</strong>, or <strong>decrease</strong>. The estimation is calculate by whether the number of check_ins increase or decrease, and the weight is <strong>1.1</strong>, <strong>1</strong>, and <strong>0.9</strong>.</p>\n\n<p>But sadly, the result did not improve. I guess the reason is because KNN already take the year parameter into account, and give recent year a closer distance.</p>\n\n<p>I think the idea of predicting diminish place_ids should intuitively work, but I am not sure how to better include this information into the final model.</p>\n\n<p>That's all my thoughts and findings.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 126016,
      "author_name": "zeroblue",
      "author_url": "",
      "post_date": "07/05/2016 14:57:51",
      "content": "<p>I also tried this, using various forms of regression to identify places that were declining in popularity, but like you, my results did not improve.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 126305,
      "author_name": "kirkir",
      "author_url": "",
      "post_date": "07/07/2016 15:52:01",
      "content": "<p>There are ~500 place_ids that showed up only during the 1st half of 1st year, while the test set is for the 2nd half of 2nd year. Deleting these got 0.00001 change in LB score. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "125964": "As title.\r\n\r\nMight be too late to ask. But since no one ask this question, I just want to ask if anyone did this, and whether it could improve LB score.\r\n\r\nIf any similar work related to this, I remember someone claim a slight increase if randomly delete some of the most popular place_ids.",
    "125966": "This idea is interesting. I am looking forward to hear any insight.",
    "125967": "Li Li, Thanks for replying.\r\n\r\nI have tried to multiply a weight to the final probability generated by KNN before taking the top 3 ranks. The weight is just a rough estimate of whether the place id will **increase**, **remain the same level**, or **decrease**. The estimation is calculate by whether the number of check_ins increase or decrease, and the weight is **1.1**, **1**, and **0.9**.\r\n\r\nBut sadly, the result did not improve. I guess the reason is because KNN already take the year parameter into account, and give recent year a closer distance.\r\n\r\nI think the idea of predicting diminish place_ids should intuitively work, but I am not sure how to better include this information into the final model.\r\n\r\nThat's all my thoughts and findings.",
    "126016": "I also tried this, using various forms of regression to identify places that were declining in popularity, but like you, my results did not improve.",
    "126305": "There are ~500 place_ids that showed up only during the 1st half of 1st year, while the test set is for the 2nd half of 2nd year. Deleting these got 0.00001 change in LB score."
  },
  "source": "meta"
}