{
  "id": 56349,
  "title": "Solution #55 up on github",
  "url": "/competitions/talkingdata-adtracking-fraud-detection/writeups/stys-solution-55-up-on-github",
  "author_name": "",
  "post_date": "2018-05-08T21:52:19.767301700Z",
  "votes": 13,
  "comment_count": 1,
  "views": 0,
  "content": "<p>This was an awesome competition! Thanks to the organizers and to all kagglers for sharing ideas and kernels! Last day was crazy, just three hours before the deadline I was kicked out of top 100 on public leaderboard. Surprisingly my submissions scored much better on the private leaderboard, climbing up almost 70 positions. Here is what I used:\n- Aggregate factors from Baris' kernel computed on all data with test_supplement\n- Prev/next click times (1-, 2-, 3-step) computed on all data with test_supplement\n- LIBFFM out-of-fold using 5 folds\n- LGBM on 50% data (using all is_attributed=1 and 50% of is_attributed=0)\n- Average of a few LGBM models trained on different subsets of data and factors for the final submission\n- Training with 96GB RAM\n- Store data column-wise in binary formats for fast loading\n- HOCON configurations</p>\n\n<p>My code for this competition is on github <a href=\"https://github.com/stys/kaggle-talkingdata-adtracking-fraud-detection\">https://github.com/stys/kaggle-talkingdata-adtracking-fraud-detection</a></p>",
  "messages": [
    {
      "id": "325803",
      "postDate": "05/08/2018 21:52:19",
      "content": "<p>This was an awesome competition! Thanks to the organizers and to all kagglers for sharing ideas and kernels! Last day was crazy, just three hours before the deadline I was kicked out of top 100 on public leaderboard. Surprisingly my submissions scored much better on the private leaderboard, climbing up almost 70 positions. Here is what I used:\n- Aggregate factors from Baris' kernel computed on all data with test_supplement\n- Prev/next click times (1-, 2-, 3-step) computed on all data with test_supplement\n- LIBFFM out-of-fold using 5 folds\n- LGBM on 50% data (using all is_attributed=1 and 50% of is_attributed=0)\n- Average of a few LGBM models trained on different subsets of data and factors for the final submission\n- Training with 96GB RAM\n- Store data column-wise in binary formats for fast loading\n- HOCON configurations</p>\n\n<p>My code for this competition is on github <a href=\"https://github.com/stys/kaggle-talkingdata-adtracking-fraud-detection\">https://github.com/stys/kaggle-talkingdata-adtracking-fraud-detection</a></p>",
      "rawMarkdown": "This was an awesome competition! Thanks to the organizers and to all kagglers for sharing ideas and kernels! Last day was crazy, just three hours before the deadline I was kicked out of top 100 on public leaderboard. Surprisingly my submissions scored much better on the private leaderboard, climbing up almost 70 positions. Here is what I used:\n- Aggregate factors from Baris' kernel computed on all data with test_supplement\n- Prev/next click times (1-, 2-, 3-step) computed on all data with test_supplement\n- LIBFFM out-of-fold using 5 folds\n- LGBM on 50% data (using all is_attributed=1 and 50% of is_attributed=0)\n- Average of a few LGBM models trained on different subsets of data and factors for the final submission\n- Training with 96GB RAM\n- Store data column-wise in binary formats for fast loading\n- HOCON configurations\n\nMy code for this competition is on github https://github.com/stys/kaggle-talkingdata-adtracking-fraud-detection",
      "votes": null
    },
    {
      "id": "326579",
      "postDate": "05/10/2018 01:19:56",
      "content": "<p>Thank you for sharing your repository with nice documents. It's very informative and helpful if solutions are provided with its repository.</p>",
      "rawMarkdown": "Thank you for sharing your repository with nice documents. It's very informative and helpful if solutions are provided with its repository.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 326579,
      "author_name": "fujihiro",
      "author_url": "",
      "post_date": "05/10/2018 01:19:56",
      "content": "<p>Thank you for sharing your repository with nice documents. It's very informative and helpful if solutions are provided with its repository.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "325803": "This was an awesome competition! Thanks to the organizers and to all kagglers for sharing ideas and kernels! Last day was crazy, just three hours before the deadline I was kicked out of top 100 on public leaderboard. Surprisingly my submissions scored much better on the private leaderboard, climbing up almost 70 positions. Here is what I used:\n- Aggregate factors from Baris' kernel computed on all data with test_supplement\n- Prev/next click times (1-, 2-, 3-step) computed on all data with test_supplement\n- LIBFFM out-of-fold using 5 folds\n- LGBM on 50% data (using all is_attributed=1 and 50% of is_attributed=0)\n- Average of a few LGBM models trained on different subsets of data and factors for the final submission\n- Training with 96GB RAM\n- Store data column-wise in binary formats for fast loading\n- HOCON configurations\n\nMy code for this competition is on github https://github.com/stys/kaggle-talkingdata-adtracking-fraud-detection",
    "326579": "Thank you for sharing your repository with nice documents. It's very informative and helpful if solutions are provided with its repository."
  },
  "source": "meta"
}