{
  "id": 195836,
  "title": "Has anyone inspected what tags mean?",
  "url": "/competitions/riiid-test-answer-prediction/discussion/195836",
  "author_name": "",
  "post_date": "2020-11-07T19:42:01.653131400Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I suspected that question and lecture tags, when equal, mean they are on the same topic. There is less lecture tags than question tags, but they have alot of equal tags. So if someone had seen this tag they might answer correctly. But I'm still going to look at this. </p>\n<p>Has anyone done some research on it?</p>",
  "messages": [
    {
      "id": "1072089",
      "postDate": "11/07/2020 19:42:01",
      "content": "<p>I suspected that question and lecture tags, when equal, mean they are on the same topic. There is less lecture tags than question tags, but they have alot of equal tags. So if someone had seen this tag they might answer correctly. But I'm still going to look at this. </p>\n<p>Has anyone done some research on it?</p>",
      "rawMarkdown": "I suspected that question and lecture tags, when equal, mean they are on the same topic. There is less lecture tags than question tags, but they have alot of equal tags. So if someone had seen this tag they might answer correctly. But I'm still going to look at this. \n\nHas anyone done some research on it?",
      "votes": null
    },
    {
      "id": "1073701",
      "postDate": "11/09/2020 20:26:08",
      "content": "<p>Hello, I haven't investigated, but I've used them to feed an embedding layer. The attachment is the result of applying t-SNE to the trained tag embeddings. I saw this post and decided to plot them for you. Also, after plotting them I've been thinking about if they are really usefull or not, too much sparsity.</p>",
      "rawMarkdown": "Hello, I haven't investigated, but I've used them to feed an embedding layer. The attachment is the result of applying t-SNE to the trained tag embeddings. I saw this post and decided to plot them for you. Also, after plotting them I've been thinking about if they are really usefull or not, too much sparsity.",
      "votes": null
    },
    {
      "id": "1073708",
      "postDate": "11/09/2020 20:35:57",
      "content": "<p>I used the tags as categorical features in my lgbm model. The first tags of the questions were the only ones that gave some real improvment, and it gives alot of gain, more than getting the accuracy for the tags. Coincidentally, the first tag has the most matching lecture tags.  So it's a good feature but I still don't know why.</p>",
      "rawMarkdown": "I used the tags as categorical features in my lgbm model. The first tags of the questions were the only ones that gave some real improvment, and it gives alot of gain, more than getting the accuracy for the tags. Coincidentally, the first tag has the most matching lecture tags.  So it's a good feature but I still don't know why.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1073701,
      "author_name": "claverru",
      "author_url": "",
      "post_date": "11/09/2020 20:26:08",
      "content": "<p>Hello, I haven't investigated, but I've used them to feed an embedding layer. The attachment is the result of applying t-SNE to the trained tag embeddings. I saw this post and decided to plot them for you. Also, after plotting them I've been thinking about if they are really usefull or not, too much sparsity.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1073708,
          "author_name": "iuryck",
          "author_url": "",
          "post_date": "11/09/2020 20:35:57",
          "content": "<p>I used the tags as categorical features in my lgbm model. The first tags of the questions were the only ones that gave some real improvment, and it gives alot of gain, more than getting the accuracy for the tags. Coincidentally, the first tag has the most matching lecture tags.  So it's a good feature but I still don't know why.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1072089": "I suspected that question and lecture tags, when equal, mean they are on the same topic. There is less lecture tags than question tags, but they have alot of equal tags. So if someone had seen this tag they might answer correctly. But I'm still going to look at this. \n\nHas anyone done some research on it?",
    "1073701": "Hello, I haven't investigated, but I've used them to feed an embedding layer. The attachment is the result of applying t-SNE to the trained tag embeddings. I saw this post and decided to plot them for you. Also, after plotting them I've been thinking about if they are really usefull or not, too much sparsity.",
    "1073708": "I used the tags as categorical features in my lgbm model. The first tags of the questions were the only ones that gave some real improvment, and it gives alot of gain, more than getting the accuracy for the tags. Coincidentally, the first tag has the most matching lecture tags.  So it's a good feature but I still don't know why."
  },
  "source": "meta"
}