{
  "id": 197648,
  "title": "Why the LGBM approach?",
  "url": "/competitions/riiid-test-answer-prediction/discussion/197648",
  "author_name": "",
  "post_date": "2020-11-17T12:03:24.101434700Z",
  "votes": 8,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Why are most contestants using LGBMs? Am I missing something here?</p>",
  "messages": [
    {
      "id": "1081905",
      "postDate": "11/17/2020 12:03:24",
      "content": "<p>Why are most contestants using LGBMs? Am I missing something here?</p>",
      "rawMarkdown": "Why are most contestants using LGBMs? Am I missing something here?",
      "votes": null
    },
    {
      "id": "1086337",
      "postDate": "11/21/2020 15:49:50",
      "content": "<p>What I can think is, because it’s a code competition with limitation that the code can’t run beyond 9 hours.  LGBM is one of the fastest algorithm.</p>",
      "rawMarkdown": "What I can think is, because it’s a code competition with limitation that the code can’t run beyond 9 hours.  LGBM is one of the fastest algorithm.",
      "votes": null
    },
    {
      "id": "1086340",
      "postDate": "11/21/2020 15:54:03",
      "content": "<p>I had the same question in my mind , but as mentioned by previous post , LGBM is more suitable for huge datasets , one of things I also was thinking was why isn't there usage of NNs in the competition but perhaps it would lead to creation of a huge network which might be way too complex…</p>",
      "rawMarkdown": "I had the same question in my mind , but as mentioned by previous post , LGBM is more suitable for huge datasets , one of things I also was thinking was why isn't there usage of NNs in the competition but perhaps it would lead to creation of a huge network which might be way too complex...",
      "votes": null
    },
    {
      "id": "1086514",
      "postDate": "11/21/2020 18:39:14",
      "content": "<p>Well, people and the organisers have used Transformer's, it'c complicated but not impossible to use it here. ( PS Check other threads)</p>",
      "rawMarkdown": "Well, people and the organisers have used Transformer's, it'c complicated but not impossible to use it here. ( PS Check other threads)",
      "votes": null
    },
    {
      "id": "1086640",
      "postDate": "11/21/2020 22:05:35",
      "content": "<p>You can use pre-trained models. Which means you can code and train your model elsewhere. Only the inference should be under 9 hours.</p>",
      "rawMarkdown": "You can use pre-trained models. Which means you can code and train your model elsewhere. Only the inference should be under 9 hours.",
      "votes": null
    },
    {
      "id": "1086939",
      "postDate": "11/22/2020 07:39:20",
      "content": "<p>Using lgbm, catboost and keras NN - local CV for all three are very similar on the same dataset of features.  Not really tried to compare performance with regards to speed on kaggle.</p>\n<p>I suspect I will end up with focus on the NN mostly because I feel more confident on tweaking for over-fit with tensorflow/keras than I do with lgbm or catboost.  </p>",
      "rawMarkdown": "Using lgbm, catboost and keras NN - local CV for all three are very similar on the same dataset of features.  Not really tried to compare performance with regards to speed on kaggle.\n\nI suspect I will end up with focus on the NN mostly because I feel more confident on tweaking for over-fit with tensorflow/keras than I do with lgbm or catboost.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1086337,
      "author_name": "dwin183287",
      "author_url": "",
      "post_date": "11/21/2020 15:49:50",
      "content": "<p>What I can think is, because it’s a code competition with limitation that the code can’t run beyond 9 hours.  LGBM is one of the fastest algorithm.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1086640,
          "author_name": "kosmos",
          "author_url": "",
          "post_date": "11/21/2020 22:05:35",
          "content": "<p>You can use pre-trained models. Which means you can code and train your model elsewhere. Only the inference should be under 9 hours.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1086340,
      "author_name": "sahilmaheshwari",
      "author_url": "",
      "post_date": "11/21/2020 15:54:03",
      "content": "<p>I had the same question in my mind , but as mentioned by previous post , LGBM is more suitable for huge datasets , one of things I also was thinking was why isn't there usage of NNs in the competition but perhaps it would lead to creation of a huge network which might be way too complex…</p>",
      "votes": null,
      "replies": [
        {
          "id": 1086514,
          "author_name": "adityaecdrid",
          "author_url": "",
          "post_date": "11/21/2020 18:39:14",
          "content": "<p>Well, people and the organisers have used Transformer's, it'c complicated but not impossible to use it here. ( PS Check other threads)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1086939,
      "author_name": "pcjimmmy",
      "author_url": "",
      "post_date": "11/22/2020 07:39:20",
      "content": "<p>Using lgbm, catboost and keras NN - local CV for all three are very similar on the same dataset of features.  Not really tried to compare performance with regards to speed on kaggle.</p>\n<p>I suspect I will end up with focus on the NN mostly because I feel more confident on tweaking for over-fit with tensorflow/keras than I do with lgbm or catboost.  </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1081905": "Why are most contestants using LGBMs? Am I missing something here?",
    "1086337": "What I can think is, because it’s a code competition with limitation that the code can’t run beyond 9 hours.  LGBM is one of the fastest algorithm.",
    "1086340": "I had the same question in my mind , but as mentioned by previous post , LGBM is more suitable for huge datasets , one of things I also was thinking was why isn't there usage of NNs in the competition but perhaps it would lead to creation of a huge network which might be way too complex...",
    "1086514": "Well, people and the organisers have used Transformer's, it'c complicated but not impossible to use it here. ( PS Check other threads)",
    "1086640": "You can use pre-trained models. Which means you can code and train your model elsewhere. Only the inference should be under 9 hours.",
    "1086939": "Using lgbm, catboost and keras NN - local CV for all three are very similar on the same dataset of features.  Not really tried to compare performance with regards to speed on kaggle.\n\nI suspect I will end up with focus on the NN mostly because I feel more confident on tweaking for over-fit with tensorflow/keras than I do with lgbm or catboost."
  },
  "source": "meta"
}