{
  "id": 79430,
  "title": "How high can you go without Deep Learning?",
  "url": "/competitions/vsb-power-line-fault-detection/discussion/79430",
  "author_name": "",
  "post_date": "2019-02-03T23:40:24.021525800Z",
  "votes": 3,
  "comment_count": 10,
  "views": 0,
  "content": "<p>Hey all,</p>\n\n<p>Just curious : how far have you guys gotten without DL models?</p>\n\n<p>I'm at 0.641 using LGBM, but my creativity is running out and I'm wondering if there is still a large room for improvement before pulling out the heavy DL guns?</p>\n\n<p>Cheers</p>\n\n<p>(update: 0.652 now, let's see)</p>",
  "messages": [
    {
      "id": "465757",
      "postDate": "02/03/2019 23:40:24",
      "content": "<p>Hey all,</p>\n\n<p>Just curious : how far have you guys gotten without DL models?</p>\n\n<p>I'm at 0.641 using LGBM, but my creativity is running out and I'm wondering if there is still a large room for improvement before pulling out the heavy DL guns?</p>\n\n<p>Cheers</p>\n\n<p>(update: 0.652 now, let's see)</p>",
      "rawMarkdown": "Hey all,\n\nJust curious : how far have you guys gotten without DL models?\n\nI'm at 0.641 using LGBM, but my creativity is running out and I'm wondering if there is still a large room for improvement before pulling out the heavy DL guns?\n\nCheers\n\n(update: 0.652 now, let's see)",
      "votes": null
    },
    {
      "id": "466144",
      "postDate": "02/04/2019 19:06:02",
      "content": "<p>0.641 looks great! I got 0.620 with LGBM and struggled for two weeks before finally giving in to DL.</p>",
      "rawMarkdown": "0.641 looks great! I got 0.620 with LGBM and struggled for two weeks before finally giving in to DL.",
      "votes": null
    },
    {
      "id": "466145",
      "postDate": "02/04/2019 19:09:34",
      "content": "<p>Sorry for such simple question, but what does \"DL\" mean?</p>",
      "rawMarkdown": "Sorry for such simple question, but what does \"DL\" mean?",
      "votes": null
    },
    {
      "id": "466160",
      "postDate": "02/04/2019 19:39:05",
      "content": "<p>DL means Deep Learning. A catch all term for various types of artificial neural networks or multilayer perceptron (MLP).</p>",
      "rawMarkdown": "DL means Deep Learning. A catch all term for various types of artificial neural networks or multilayer perceptron (MLP).",
      "votes": null
    },
    {
      "id": "466182",
      "postDate": "02/04/2019 20:30:36",
      "content": "<p>@Jack  Thanks, I just didn't recognize the abbreviation.</p>",
      "rawMarkdown": "Jack  Thanks, I just didn't recognize the abbreviation.",
      "votes": null
    },
    {
      "id": "466209",
      "postDate": "02/04/2019 21:29:37",
      "content": "<p><a href=\"/dslate\">@dslate</a> No problem, my second sentence was a bit unnecessary 😄</p>",
      "rawMarkdown": "dslate No problem, my second sentence was a bit unnecessary 😄",
      "votes": null
    },
    {
      "id": "466228",
      "postDate": "02/04/2019 22:27:55",
      "content": "<p>I feel you, guess I'm giving in too- thanks for sharing!</p>",
      "rawMarkdown": "I feel you, guess I'm giving in too- thanks for sharing!",
      "votes": null
    },
    {
      "id": "467578",
      "postDate": "02/07/2019 11:09:03",
      "content": "<p>Why is everyone so eager to use neural networks on this problem?  There are only about 20k rows in the test set.  Also fairly simple statistics seem to provide good baselines that can be ensembled.  Noisy data with tons of feature columns seems more of a bias vs variance problem than a DL problem. </p>",
      "rawMarkdown": "Why is everyone so eager to use neural networks on this problem?  There are only about 20k rows in the test set.  Also fairly simple statistics seem to provide good baselines that can be ensembled.  Noisy data with tons of feature columns seems more of a bias vs variance problem than a DL problem.",
      "votes": null
    },
    {
      "id": "468236",
      "postDate": "02/08/2019 14:49:28",
      "content": "<p>Because it seems to be working better than feature-based approaches. I guess we don't have enough domain knowledge to be able to extract strong features.</p>",
      "rawMarkdown": "Because it seems to be working better than feature-based approaches. I guess we don't have enough domain knowledge to be able to extract strong features.",
      "votes": null
    },
    {
      "id": "473779",
      "postDate": "02/18/2019 13:53:17",
      "content": "<p>How about your local cv?</p>",
      "rawMarkdown": "How about your local cv?",
      "votes": null
    },
    {
      "id": "473951",
      "postDate": "02/18/2019 18:35:45",
      "content": "<p>CV surprisingly reliable ~0.69x for LB 0.65x</p>",
      "rawMarkdown": "CV surprisingly reliable ~0.69x for LB 0.65x",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 466144,
      "author_name": "sangxia",
      "author_url": "",
      "post_date": "02/04/2019 19:06:02",
      "content": "<p>0.641 looks great! I got 0.620 with LGBM and struggled for two weeks before finally giving in to DL.</p>",
      "votes": null,
      "replies": [
        {
          "id": 466228,
          "author_name": "ganfear",
          "author_url": "",
          "post_date": "02/04/2019 22:27:55",
          "content": "<p>I feel you, guess I'm giving in too- thanks for sharing!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 466145,
      "author_name": "dslate",
      "author_url": "",
      "post_date": "02/04/2019 19:09:34",
      "content": "<p>Sorry for such simple question, but what does \"DL\" mean?</p>",
      "votes": null,
      "replies": [
        {
          "id": 466160,
          "author_name": "jackvial",
          "author_url": "",
          "post_date": "02/04/2019 19:39:05",
          "content": "<p>DL means Deep Learning. A catch all term for various types of artificial neural networks or multilayer perceptron (MLP).</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 466182,
          "author_name": "dslate",
          "author_url": "",
          "post_date": "02/04/2019 20:30:36",
          "content": "<p>@Jack  Thanks, I just didn't recognize the abbreviation.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 466209,
          "author_name": "jackvial",
          "author_url": "",
          "post_date": "02/04/2019 21:29:37",
          "content": "<p><a href=\"/dslate\">@dslate</a> No problem, my second sentence was a bit unnecessary 😄</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 467578,
      "author_name": "troymann",
      "author_url": "",
      "post_date": "02/07/2019 11:09:03",
      "content": "<p>Why is everyone so eager to use neural networks on this problem?  There are only about 20k rows in the test set.  Also fairly simple statistics seem to provide good baselines that can be ensembled.  Noisy data with tons of feature columns seems more of a bias vs variance problem than a DL problem. </p>",
      "votes": null,
      "replies": [
        {
          "id": 468236,
          "author_name": "maxhalford",
          "author_url": "",
          "post_date": "02/08/2019 14:49:28",
          "content": "<p>Because it seems to be working better than feature-based approaches. I guess we don't have enough domain knowledge to be able to extract strong features.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 473779,
      "author_name": "blackboards",
      "author_url": "",
      "post_date": "02/18/2019 13:53:17",
      "content": "<p>How about your local cv?</p>",
      "votes": null,
      "replies": [
        {
          "id": 473951,
          "author_name": "ganfear",
          "author_url": "",
          "post_date": "02/18/2019 18:35:45",
          "content": "<p>CV surprisingly reliable ~0.69x for LB 0.65x</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "465757": "Hey all,\n\nJust curious : how far have you guys gotten without DL models?\n\nI'm at 0.641 using LGBM, but my creativity is running out and I'm wondering if there is still a large room for improvement before pulling out the heavy DL guns?\n\nCheers\n\n(update: 0.652 now, let's see)",
    "466144": "0.641 looks great! I got 0.620 with LGBM and struggled for two weeks before finally giving in to DL.",
    "466145": "Sorry for such simple question, but what does \"DL\" mean?",
    "466160": "DL means Deep Learning. A catch all term for various types of artificial neural networks or multilayer perceptron (MLP).",
    "466182": "Jack  Thanks, I just didn't recognize the abbreviation.",
    "466209": "dslate No problem, my second sentence was a bit unnecessary 😄",
    "466228": "I feel you, guess I'm giving in too- thanks for sharing!",
    "467578": "Why is everyone so eager to use neural networks on this problem?  There are only about 20k rows in the test set.  Also fairly simple statistics seem to provide good baselines that can be ensembled.  Noisy data with tons of feature columns seems more of a bias vs variance problem than a DL problem.",
    "468236": "Because it seems to be working better than feature-based approaches. I guess we don't have enough domain knowledge to be able to extract strong features.",
    "473779": "How about your local cv?",
    "473951": "CV surprisingly reliable ~0.69x for LB 0.65x"
  },
  "source": "meta"
}