{
  "id": 336415,
  "title": "9th place solution",
  "content": "<h2>setting</h2>\n<p>evaluate local by 2fold CV<br>\ngroupKfold by POI</p>\n<h2>1st : candidate selection</h2>\n<ul>\n<li>selected 400 candidate for each id<ul>\n<li>knn with latlon: 350 candidate</li>\n<li>knn with latlon + name tfidf: 50 candidate</li></ul></li>\n<li>Max IoU: 0.994, 420M candidates</li>\n</ul>\n<h2>2nd : modelling light model</h2>\n<ul>\n<li>catboost with tfidf similarity <ul>\n<li>features (40 features)<ul>\n<li>text similarity with tfidf</li>\n<li>raw &amp; diff int of text</li>\n<li>raw &amp; diff latlon</li></ul></li>\n<li>threshold is 0.005</li>\n<li>Max IoU: 0.987, 3M candidates</li></ul></li>\n</ul>\n<h2>3rd : modelling</h2>\n<ul>\n<li>catboost and lightgbm<ul>\n<li>catboost with text encoding each text<ul>\n<li>0.010 improved with text encoding</li></ul></li>\n<li>features (2nd40 + 100 features)</li>\n<li>text similarities with gesh, LCS, leven, jaccard… 12 methods</li>\n<li>count ids, match ids</li>\n<li>target encoding each pair of ['name', 'name_match'], <a href=\"2fold\" target=\"_blank\">'categories', 'categories_match'</a></li></ul></li>\n</ul>\n<h2>4th : post_process</h2>\n<ul>\n<li>calc the node predictions with avg of neighbors predictions, order by predictions<ul>\n<li>thresholds: 0.4</li></ul></li>\n</ul>\n<h2>score</h2>\n<ul>\n<li>Final CV: 0.905, 2fold CV</li>\n<li>LB: <ul>\n<li>catboost(cb): 0.925</li>\n<li>cb + post_process(pp): 0.930</li>\n<li>cb + lgb + pp: 0.933</li>\n<li>cb + lgb + pp + training all data(1fold): 0.943</li>\n<li>cb + lgb + pp + all data + ensemble 5models  0.945</li>\n<li>cb + lgb + pp + all data + ensemble + increase iterations(over fit): 0.948</li></ul></li>\n</ul>",
  "comments": [
    {
      "id": 1854136,
      "author_name": "Naman Sharma",
      "author_url": "",
      "post_date": "2022-07-13T13:04:14.657000",
      "content": "<p>Really Commendable work !! thanksss for sharinggg👌</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1853536,
      "author_name": "The Devastator",
      "author_url": "",
      "post_date": "2022-07-13T00:12:38.650000",
      "content": "<p>Nice job on your 9th place finish! Your solution is very informative and helpful. <br>\nThank you for sharing!</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1853560,
          "author_name": "taksai",
          "author_url": "",
          "post_date": "2022-07-13T00:42:14.873000",
          "content": "<p>Thank you!!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1853054,
      "author_name": "pao",
      "author_url": "",
      "post_date": "2022-07-12T14:57:15.840000",
      "content": "<p>Thank you for sharing, and it is interesting that you will get such high score by only GBDT.</p>\n<p>About 1st and 2nd stage, I think it is required huge memories because 400 candidates and 40 features are very big. (Our team also use first simple model, but the size is smaller than yours)<br>\nHow much machine resource do you use in train? and how long is it spend in inference phase?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1853558,
          "author_name": "taksai",
          "author_url": "",
          "post_date": "2022-07-13T00:37:00.913000",
          "content": "<p>Thank you for comment. And, congratulations for winning!!<br>\nI use memory about 500GB~1TB in train, 2~3hours spent in 2nd inference phase (including make features).</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 1853579,
          "author_name": "pao",
          "author_url": "",
          "post_date": "2022-07-13T01:28:29.470000",
          "content": "<p>Wow, very huge machine….!!!<br>\nThank you for telling!!</p>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "comment_count": 6,
  "messages": [
    {
      "id": 1851120,
      "postDate": "2022-07-11T03:21:32.453Z",
      "content": "<h2>setting</h2>\n<p>evaluate local by 2fold CV<br>\ngroupKfold by POI</p>\n<h2>1st : candidate selection</h2>\n<ul>\n<li>selected 400 candidate for each id<ul>\n<li>knn with latlon: 350 candidate</li>\n<li>knn with latlon + name tfidf: 50 candidate</li></ul></li>\n<li>Max IoU: 0.994, 420M candidates</li>\n</ul>\n<h2>2nd : modelling light model</h2>\n<ul>\n<li>catboost with tfidf similarity <ul>\n<li>features (40 features)<ul>\n<li>text similarity with tfidf</li>\n<li>raw &amp; diff int of text</li>\n<li>raw &amp; diff latlon</li></ul></li>\n<li>threshold is 0.005</li>\n<li>Max IoU: 0.987, 3M candidates</li></ul></li>\n</ul>\n<h2>3rd : modelling</h2>\n<ul>\n<li>catboost and lightgbm<ul>\n<li>catboost with text encoding each text<ul>\n<li>0.010 improved with text encoding</li></ul></li>\n<li>features (2nd40 + 100 features)</li>\n<li>text similarities with gesh, LCS, leven, jaccard… 12 methods</li>\n<li>count ids, match ids</li>\n<li>target encoding each pair of ['name', 'name_match'], <a href=\"2fold\" target=\"_blank\">'categories', 'categories_match'</a></li></ul></li>\n</ul>\n<h2>4th : post_process</h2>\n<ul>\n<li>calc the node predictions with avg of neighbors predictions, order by predictions<ul>\n<li>thresholds: 0.4</li></ul></li>\n</ul>\n<h2>score</h2>\n<ul>\n<li>Final CV: 0.905, 2fold CV</li>\n<li>LB: <ul>\n<li>catboost(cb): 0.925</li>\n<li>cb + post_process(pp): 0.930</li>\n<li>cb + lgb + pp: 0.933</li>\n<li>cb + lgb + pp + training all data(1fold): 0.943</li>\n<li>cb + lgb + pp + all data + ensemble 5models  0.945</li>\n<li>cb + lgb + pp + all data + ensemble + increase iterations(over fit): 0.948</li></ul></li>\n</ul>",
      "rawMarkdown": "## setting\nevaluate local by 2fold CV\ngroupKfold by POI\n\n## 1st : candidate selection\n- selected 400 candidate for each id\n  - knn with latlon: 350 candidate\n  - knn with latlon + name tfidf: 50 candidate\n-  Max IoU: 0.994, 420M candidates\n\n## 2nd : modelling light model\n- catboost with tfidf similarity \n  - features (40 features)\n     - text similarity with tfidf\n     - raw & diff int of text\n     - raw & diff latlon\n  - threshold is 0.005\n  - Max IoU: 0.987, 3M candidates\n\n## 3rd : modelling\n- catboost and lightgbm\n    - catboost with text encoding each text\n        - 0.010 improved with text encoding\n  - features (2nd40 + 100 features)\n    - text similarities with gesh, LCS, leven, jaccard... 12 methods\n    - count ids, match ids\n    - target encoding each pair of ['name', 'name_match'], ['categories', 'categories_match'] (2fold)\n\n## 4th : post_process\n- calc the node predictions with avg of neighbors predictions, order by predictions\n    - thresholds: 0.4\n\n## score\n- Final CV: 0.905, 2fold CV\n- LB: \n    - catboost(cb): 0.925\n    - cb + post_process(pp): 0.930\n    - cb + lgb + pp: 0.933\n    - cb + lgb + pp + training all data(1fold): 0.943\n    - cb + lgb + pp + all data + ensemble 5models  0.945\n    - cb + lgb + pp + all data + ensemble + increase iterations(over fit): 0.948",
      "votes": 32
    },
    {
      "id": 1854136,
      "postDate": "2022-07-13T13:04:14.657Z",
      "content": "<p>Really Commendable work !! thanksss for sharinggg👌</p>",
      "rawMarkdown": "Really Commendable work !! thanksss for sharinggg👌",
      "votes": 1
    },
    {
      "id": 1853536,
      "postDate": "2022-07-13T00:12:38.650Z",
      "content": "<p>Nice job on your 9th place finish! Your solution is very informative and helpful. <br>\nThank you for sharing!</p>",
      "rawMarkdown": "Nice job on your 9th place finish! Your solution is very informative and helpful. \nThank you for sharing!\n",
      "votes": 1,
      "replies": [
        {
          "id": 1853560,
          "postDate": "2022-07-13T00:42:14.873Z",
          "content": "<p>Thank you!!</p>",
          "rawMarkdown": "Thank you!!"
        }
      ]
    },
    {
      "id": 1853054,
      "postDate": "2022-07-12T14:57:15.840Z",
      "content": "<p>Thank you for sharing, and it is interesting that you will get such high score by only GBDT.</p>\n<p>About 1st and 2nd stage, I think it is required huge memories because 400 candidates and 40 features are very big. (Our team also use first simple model, but the size is smaller than yours)<br>\nHow much machine resource do you use in train? and how long is it spend in inference phase?</p>",
      "rawMarkdown": "Thank you for sharing, and it is interesting that you will get such high score by only GBDT.\n\nAbout 1st and 2nd stage, I think it is required huge memories because 400 candidates and 40 features are very big. (Our team also use first simple model, but the size is smaller than yours)\nHow much machine resource do you use in train? and how long is it spend in inference phase?",
      "votes": 1,
      "replies": [
        {
          "id": 1853558,
          "postDate": "2022-07-13T00:37:00.913Z",
          "content": "<p>Thank you for comment. And, congratulations for winning!!<br>\nI use memory about 500GB~1TB in train, 2~3hours spent in 2nd inference phase (including make features).</p>",
          "rawMarkdown": "Thank you for comment. And, congratulations for winning!!\nI use memory about 500GB~1TB in train, 2~3hours spent in 2nd inference phase (including make features).",
          "votes": 2
        },
        {
          "id": 1853579,
          "postDate": "2022-07-13T01:28:29.470Z",
          "content": "<p>Wow, very huge machine….!!!<br>\nThank you for telling!!</p>",
          "rawMarkdown": "Wow, very huge machine....!!!\nThank you for telling!!",
          "votes": 1
        }
      ]
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-22T08:26:26Z",
  "competition_id": 35476,
  "competition_slug": "foursquare-location-matching",
  "team_id": 8638370,
  "solution_url": "/competitions/foursquare-location-matching/writeups/taksai-9th-place-solution"
}