{
  "id": 76411,
  "title": "What kind of improvements we can do to existing the model?",
  "url": "/competitions/quora-insincere-questions-classification/discussion/76411",
  "author_name": "",
  "post_date": "2019-01-02T15:15:51.144445Z",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I tried to improve my kernel many times, and for the last 40-ty times I couldn't do this. So I have a question: if we have a kernel with a single model and we don't want to change this model, maybe small things, how we would try to improve score?\nMy ideas are:\n1) Text pre-processing (it helped me to improve my score on 0.003)\n2) Playing with embeddings (but it didn't improve my score, maybe someone can give an advice?)\n3) Adding layers to the main model (capsule improve my score on 0.001, batch normalization, additional dropout,  hidden size changing, additional linear layers didn't improve my score)\n4) Post-processing (unfortunately I don't know from where to start)</p>\n\n<p>I'll be very grateful for your advice.</p>",
  "messages": [
    {
      "id": "449039",
      "postDate": "01/02/2019 15:15:51",
      "content": "<p>I tried to improve my kernel many times, and for the last 40-ty times I couldn't do this. So I have a question: if we have a kernel with a single model and we don't want to change this model, maybe small things, how we would try to improve score?\nMy ideas are:\n1) Text pre-processing (it helped me to improve my score on 0.003)\n2) Playing with embeddings (but it didn't improve my score, maybe someone can give an advice?)\n3) Adding layers to the main model (capsule improve my score on 0.001, batch normalization, additional dropout,  hidden size changing, additional linear layers didn't improve my score)\n4) Post-processing (unfortunately I don't know from where to start)</p>\n\n<p>I'll be very grateful for your advice.</p>",
      "rawMarkdown": "I tried to improve my kernel many times, and for the last 40-ty times I couldn't do this. So I have a question: if we have a kernel with a single model and we don't want to change this model, maybe small things, how we would try to improve score?\nMy ideas are:\n1) Text pre-processing (it helped me to improve my score on 0.003)\n2) Playing with embeddings (but it didn't improve my score, maybe someone can give an advice?)\n3) Adding layers to the main model (capsule improve my score on 0.001, batch normalization, additional dropout,  hidden size changing, additional linear layers didn't improve my score)\n4) Post-processing (unfortunately I don't know from where to start)\n\nI'll be very grateful for your advice.",
      "votes": null
    },
    {
      "id": "449042",
      "postDate": "01/02/2019 15:19:03",
      "content": "<p>Also, I tried to add extra features words_vs_unique and caps_vs_length. It decreased my score, but after I changed preprocessing technics I got exactly the same score.</p>",
      "rawMarkdown": "Also, I tried to add extra features words_vs_unique and caps_vs_length. It decreased my score, but after I changed preprocessing technics I got exactly the same score.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 449042,
      "author_name": "psyfaker",
      "author_url": "",
      "post_date": "01/02/2019 15:19:03",
      "content": "<p>Also, I tried to add extra features words_vs_unique and caps_vs_length. It decreased my score, but after I changed preprocessing technics I got exactly the same score.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "449039": "I tried to improve my kernel many times, and for the last 40-ty times I couldn't do this. So I have a question: if we have a kernel with a single model and we don't want to change this model, maybe small things, how we would try to improve score?\nMy ideas are:\n1) Text pre-processing (it helped me to improve my score on 0.003)\n2) Playing with embeddings (but it didn't improve my score, maybe someone can give an advice?)\n3) Adding layers to the main model (capsule improve my score on 0.001, batch normalization, additional dropout,  hidden size changing, additional linear layers didn't improve my score)\n4) Post-processing (unfortunately I don't know from where to start)\n\nI'll be very grateful for your advice.",
    "449042": "Also, I tried to add extra features words_vs_unique and caps_vs_length. It decreased my score, but after I changed preprocessing technics I got exactly the same score."
  },
  "source": "meta"
}