{
  "id": 22142,
  "title": "Solution sharing",
  "url": "/competitions/facebook-v-predicting-check-ins/discussion/22142",
  "author_name": "Chengcen",
  "post_date": "2016-07-10T18:54:51.647000",
  "votes": 2,
  "comment_count": 0,
  "views": 258,
  "content": "<p>I really want to thanks everyone in the forum to share their methods and inspire other people. I enjoy myself in this competition and spend some time on it. My ranking(#33) is not high compared to Kaggle masters but it is my personal record. So I also want to share my solution(based on many other successful script).</p>\n\n<p>The basic ideas are:</p>\n\n<p>1.</p>\n\n<p>Two python sript based on similar methods: working with subset of the whole sample by chopping into grids</p>\n\n<p>Original code : grid_v1.py</p>\n\n<p>Improved code : grid_v2.py</p>\n\n<p>Add data for periodic time that hit the boundary: make sure 23:00 and 1:00 closer</p>\n\n<p>Add border_augment, including points near the boundary</p>\n\n<p>2.</p>\n\n<p>Initial idea to improve: based on the grid_v1. Run several models: KNN, Xgboost, with different feature included(log_accuracy, drop_day, drop accuracy), also with different grid number</p>\n\n<p>Make Ensemble using the make_ensemble.py(Based on MAP@5)</p>\n\n<p>Use gene_sub.py to make final submission. Without any parameter tuning, this can reach around 0.59</p>\n\n<p>Special thanks to YijunMa for this ensemble idea</p>\n\n<ol start=\"3\">\n<li><p>Run hyOpt.py to fine tuning the parameter</p></li>\n</ol>\n\n<p>4.</p>\n\n<p>Including the RandomForest into final prediction is helpful. Also, we should make ensemble on three models directly based on the probability. grid_v2 use combination of KNN, Xgboost, and RF. Try mix_opt.py to find the best ratio between three models. Should have more sophisticated way to figure out the weight though.</p>\n\n<p>This grid_v2.py should give the score around 0.60078</p>\n\n<p><a href=\"http://github.com/samshaq19912009/Facebook_check_in\">Github link</a></p>",
  "messages": [
    {
      "id": 126612,
      "postDate": "2016-07-10T18:54:51.647Z",
      "content": "<p>I really want to thanks everyone in the forum to share their methods and inspire other people. I enjoy myself in this competition and spend some time on it. My ranking(#33) is not high compared to Kaggle masters but it is my personal record. So I also want to share my solution(based on many other successful script).</p>\n\n<p>The basic ideas are:</p>\n\n<p>1.</p>\n\n<p>Two python sript based on similar methods: working with subset of the whole sample by chopping into grids</p>\n\n<p>Original code : grid_v1.py</p>\n\n<p>Improved code : grid_v2.py</p>\n\n<p>Add data for periodic time that hit the boundary: make sure 23:00 and 1:00 closer</p>\n\n<p>Add border_augment, including points near the boundary</p>\n\n<p>2.</p>\n\n<p>Initial idea to improve: based on the grid_v1. Run several models: KNN, Xgboost, with different feature included(log_accuracy, drop_day, drop accuracy), also with different grid number</p>\n\n<p>Make Ensemble using the make_ensemble.py(Based on MAP@5)</p>\n\n<p>Use gene_sub.py to make final submission. Without any parameter tuning, this can reach around 0.59</p>\n\n<p>Special thanks to YijunMa for this ensemble idea</p>\n\n<ol start=\"3\">\n<li><p>Run hyOpt.py to fine tuning the parameter</p></li>\n</ol>\n\n<p>4.</p>\n\n<p>Including the RandomForest into final prediction is helpful. Also, we should make ensemble on three models directly based on the probability. grid_v2 use combination of KNN, Xgboost, and RF. Try mix_opt.py to find the best ratio between three models. Should have more sophisticated way to figure out the weight though.</p>\n\n<p>This grid_v2.py should give the score around 0.60078</p>\n\n<p><a href=\"http://github.com/samshaq19912009/Facebook_check_in\">Github link</a></p>",
      "rawMarkdown": "I really want to thanks everyone in the forum to share their methods and inspire other people. I enjoy myself in this competition and spend some time on it. My ranking(#33) is not high compared to Kaggle masters but it is my personal record. So I also want to share my solution(based on many other successful script).\r\n\r\nThe basic ideas are:\r\n\r\n 1.\r\n\r\nTwo python sript based on similar methods: working with subset of the whole sample by chopping into grids\r\n\r\nOriginal code : grid_v1.py\r\n\r\nImproved code : grid_v2.py\r\n\r\nAdd data for periodic time that hit the boundary: make sure 23:00 and 1:00 closer\r\n\r\nAdd border_augment, including points near the boundary\r\n\r\n2.\r\n\r\nInitial idea to improve: based on the grid_v1. Run several models: KNN, Xgboost, with different feature included(log_accuracy, drop_day, drop accuracy), also with different grid number\r\n\r\nMake Ensemble using the make_ensemble.py(Based on MAP@5)\r\n\r\nUse gene_sub.py to make final submission. Without any parameter tuning, this can reach around 0.59\r\n\r\nSpecial thanks to YijunMa for this ensemble idea\r\n\r\n3. \r\n\r\nRun hyOpt.py to fine tuning the parameter\r\n\r\n\r\n4.\r\n\r\n\r\nIncluding the RandomForest into final prediction is helpful. Also, we should make ensemble on three models directly based on the probability. grid_v2 use combination of KNN, Xgboost, and RF. Try mix_opt.py to find the best ratio between three models. Should have more sophisticated way to figure out the weight though.\r\n\r\nThis grid_v2.py should give the score around 0.60078\r\n\r\n\r\n\r\n[Github link][1]\r\n\r\n\r\n  [1]: http://github.com/samshaq19912009/Facebook_check_in",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "126612": "I really want to thanks everyone in the forum to share their methods and inspire other people. I enjoy myself in this competition and spend some time on it. My ranking(#33) is not high compared to Kaggle masters but it is my personal record. So I also want to share my solution(based on many other successful script).\r\n\r\nThe basic ideas are:\r\n\r\n 1.\r\n\r\nTwo python sript based on similar methods: working with subset of the whole sample by chopping into grids\r\n\r\nOriginal code : grid_v1.py\r\n\r\nImproved code : grid_v2.py\r\n\r\nAdd data for periodic time that hit the boundary: make sure 23:00 and 1:00 closer\r\n\r\nAdd border_augment, including points near the boundary\r\n\r\n2.\r\n\r\nInitial idea to improve: based on the grid_v1. Run several models: KNN, Xgboost, with different feature included(log_accuracy, drop_day, drop accuracy), also with different grid number\r\n\r\nMake Ensemble using the make_ensemble.py(Based on MAP@5)\r\n\r\nUse gene_sub.py to make final submission. Without any parameter tuning, this can reach around 0.59\r\n\r\nSpecial thanks to YijunMa for this ensemble idea\r\n\r\n3. \r\n\r\nRun hyOpt.py to fine tuning the parameter\r\n\r\n\r\n4.\r\n\r\n\r\nIncluding the RandomForest into final prediction is helpful. Also, we should make ensemble on three models directly based on the probability. grid_v2 use combination of KNN, Xgboost, and RF. Try mix_opt.py to find the best ratio between three models. Should have more sophisticated way to figure out the weight though.\r\n\r\nThis grid_v2.py should give the score around 0.60078\r\n\r\n\r\n\r\n[Github link][1]\r\n\r\n\r\n  [1]: http://github.com/samshaq19912009/Facebook_check_in"
  }
}