{
  "id": 72284,
  "title": "Non feature-based",
  "url": "/competitions/PLAsTiCC-2018/discussion/72284",
  "author_name": "",
  "post_date": "2018-11-21T23:22:36.649511800Z",
  "votes": 3,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Any of the teams using models that are not feature-based (e.g. deep learning ) ? </p>",
  "messages": [
    {
      "id": "425660",
      "postDate": "11/21/2018 23:22:36",
      "content": "<p>Any of the teams using models that are not feature-based (e.g. deep learning ) ? </p>",
      "rawMarkdown": "Any of the teams using models that are not feature-based (e.g. deep learning ) ?",
      "votes": null
    },
    {
      "id": "426005",
      "postDate": "11/22/2018 12:31:09",
      "content": "<p>Feature engineering is useful for deep learning in general.  I know, this is not what people say, but reality is different.  Even for images classification, look at how much training time augmentation people do.  And how they postprocess results as well.  This is not called feature engineering but it looks similar to it to me.</p>",
      "rawMarkdown": "Feature engineering is useful for deep learning in general.  I know, this is not what people say, but reality is different.  Even for images classification, look at how much training time augmentation people do.  And how they postprocess results as well.  This is not called feature engineering but it looks similar to it to me.",
      "votes": null
    },
    {
      "id": "426019",
      "postDate": "11/22/2018 12:56:33",
      "content": "<p>I would say this is feature learning engineering, as you don't care in those cases to pass your understanding of the data to your model, but to help it learn by itself. </p>\n\n<p>I think I see your point. In the end it requires designing your model, depending on the data, which is like transferring in another space the problem of feature engineering</p>",
      "rawMarkdown": "I would say this is feature learning engineering, as you don't care in those cases to pass your understanding of the data to your model, but to help it learn by itself. \n\nI think I see your point. In the end it requires designing your model, depending on the data, which is like transferring in another space the problem of feature engineering",
      "votes": null
    },
    {
      "id": "427581",
      "postDate": "11/25/2018 19:01:41",
      "content": "<p>I feel that the size of the training set is the limiting factor for deep learning methods, but the size of the test set makes unsupervised / semi-supervised approaches with deep learning very attractive. My random experiments seem to suggest that even a crude autoencoder might be able to generate consistently useful features for a tree classifier. You have to be smart about how you construct the network though, because the nature of this task (multivariate, unevenly spaced, missings) makes typical time series architectures difficult to implement and perform.</p>",
      "rawMarkdown": "I feel that the size of the training set is the limiting factor for deep learning methods, but the size of the test set makes unsupervised / semi-supervised approaches with deep learning very attractive. My random experiments seem to suggest that even a crude autoencoder might be able to generate consistently useful features for a tree classifier. You have to be smart about how you construct the network though, because the nature of this task (multivariate, unevenly spaced, missings) makes typical time series architectures difficult to implement and perform.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 426005,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "11/22/2018 12:31:09",
      "content": "<p>Feature engineering is useful for deep learning in general.  I know, this is not what people say, but reality is different.  Even for images classification, look at how much training time augmentation people do.  And how they postprocess results as well.  This is not called feature engineering but it looks similar to it to me.</p>",
      "votes": null,
      "replies": [
        {
          "id": 426019,
          "author_name": "iprapas",
          "author_url": "",
          "post_date": "11/22/2018 12:56:33",
          "content": "<p>I would say this is feature learning engineering, as you don't care in those cases to pass your understanding of the data to your model, but to help it learn by itself. </p>\n\n<p>I think I see your point. In the end it requires designing your model, depending on the data, which is like transferring in another space the problem of feature engineering</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 427581,
      "author_name": "mithrillion",
      "author_url": "",
      "post_date": "11/25/2018 19:01:41",
      "content": "<p>I feel that the size of the training set is the limiting factor for deep learning methods, but the size of the test set makes unsupervised / semi-supervised approaches with deep learning very attractive. My random experiments seem to suggest that even a crude autoencoder might be able to generate consistently useful features for a tree classifier. You have to be smart about how you construct the network though, because the nature of this task (multivariate, unevenly spaced, missings) makes typical time series architectures difficult to implement and perform.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "425660": "Any of the teams using models that are not feature-based (e.g. deep learning ) ?",
    "426005": "Feature engineering is useful for deep learning in general.  I know, this is not what people say, but reality is different.  Even for images classification, look at how much training time augmentation people do.  And how they postprocess results as well.  This is not called feature engineering but it looks similar to it to me.",
    "426019": "I would say this is feature learning engineering, as you don't care in those cases to pass your understanding of the data to your model, but to help it learn by itself. \n\nI think I see your point. In the end it requires designing your model, depending on the data, which is like transferring in another space the problem of feature engineering",
    "427581": "I feel that the size of the training set is the limiting factor for deep learning methods, but the size of the test set makes unsupervised / semi-supervised approaches with deep learning very attractive. My random experiments seem to suggest that even a crude autoencoder might be able to generate consistently useful features for a tree classifier. You have to be smart about how you construct the network though, because the nature of this task (multivariate, unevenly spaced, missings) makes typical time series architectures difficult to implement and perform."
  },
  "source": "meta"
}