{
  "id": 171475,
  "title": "Idea: Using slope instead of predicting direct FVC values?",
  "url": "/competitions/osic-pulmonary-fibrosis-progression/discussion/171475",
  "author_name": "",
  "post_date": "2020-07-31T23:32:21.075916400Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Instead of predicting the last three FVC values, maybe we could predict the slope of the trend line. Two less values to predict and less confusing to the model since there's a lot of variation. Do you think this idea will work? Thanks in advance.</p>",
  "messages": [
    {
      "id": "953621",
      "postDate": "07/31/2020 23:32:21",
      "content": "<p>Instead of predicting the last three FVC values, maybe we could predict the slope of the trend line. Two less values to predict and less confusing to the model since there's a lot of variation. Do you think this idea will work? Thanks in advance.</p>",
      "rawMarkdown": "Instead of predicting the last three FVC values, maybe we could predict the slope of the trend line. Two less values to predict and less confusing to the model since there's a lot of variation. Do you think this idea will work? Thanks in advance.",
      "votes": null
    },
    {
      "id": "954484",
      "postDate": "08/01/2020 19:13:14",
      "content": "<p>Do you mean the slope of the trend line at the three points?  Because the trend line is doesn't have to be linear, so there won't be a single slope.</p>",
      "rawMarkdown": "Do you mean the slope of the trend line at the three points?  Because the trend line is doesn't have to be linear, so there won't be a single slope.",
      "votes": null
    },
    {
      "id": "954494",
      "postDate": "08/01/2020 19:21:48",
      "content": "<p>I meant a linear trendline.</p>",
      "rawMarkdown": "I meant a linear trendline.",
      "votes": null
    },
    {
      "id": "955607",
      "postDate": "08/02/2020 18:27:58",
      "content": "<p>If you look at the highest scoring kernel right now (using Quantile Regression), the prediction is already basically linear. </p>\n\n<p>I'm not sure what kind of benefit you might expect to obtain from just a slope prediction? A slope target would necessitate an intercept target as well, and you would still need to assign an uncertainty to it...so that's 3 values to predict, which is more than the 2 already needed. Going completely linear may also mean you're missing out on a more complex function if you manage to find good features...</p>",
      "rawMarkdown": "If you look at the highest scoring kernel right now (using Quantile Regression), the prediction is already basically linear. \n\nI'm not sure what kind of benefit you might expect to obtain from just a slope prediction? A slope target would necessitate an intercept target as well, and you would still need to assign an uncertainty to it...so that's 3 values to predict, which is more than the 2 already needed. Going completely linear may also mean you're missing out on a more complex function if you manage to find good features...",
      "votes": null
    },
    {
      "id": "958917",
      "postDate": "08/05/2020 08:06:12",
      "content": "<p>In CV it seems to me that slope is easier to predict that the value, but the percent variation is even better</p>",
      "rawMarkdown": "In CV it seems to me that slope is easier to predict that the value, but the percent variation is even better",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 954484,
      "author_name": "thebayou",
      "author_url": "",
      "post_date": "08/01/2020 19:13:14",
      "content": "<p>Do you mean the slope of the trend line at the three points?  Because the trend line is doesn't have to be linear, so there won't be a single slope.</p>",
      "votes": null,
      "replies": [
        {
          "id": 954494,
          "author_name": "unkownhihi",
          "author_url": "",
          "post_date": "08/01/2020 19:21:48",
          "content": "<p>I meant a linear trendline.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 955607,
      "author_name": "seraphimstreets",
      "author_url": "",
      "post_date": "08/02/2020 18:27:58",
      "content": "<p>If you look at the highest scoring kernel right now (using Quantile Regression), the prediction is already basically linear. </p>\n\n<p>I'm not sure what kind of benefit you might expect to obtain from just a slope prediction? A slope target would necessitate an intercept target as well, and you would still need to assign an uncertainty to it...so that's 3 values to predict, which is more than the 2 already needed. Going completely linear may also mean you're missing out on a more complex function if you manage to find good features...</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 958917,
      "author_name": "abiolatti",
      "author_url": "",
      "post_date": "08/05/2020 08:06:12",
      "content": "<p>In CV it seems to me that slope is easier to predict that the value, but the percent variation is even better</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "953621": "Instead of predicting the last three FVC values, maybe we could predict the slope of the trend line. Two less values to predict and less confusing to the model since there's a lot of variation. Do you think this idea will work? Thanks in advance.",
    "954484": "Do you mean the slope of the trend line at the three points?  Because the trend line is doesn't have to be linear, so there won't be a single slope.",
    "954494": "I meant a linear trendline.",
    "955607": "If you look at the highest scoring kernel right now (using Quantile Regression), the prediction is already basically linear. \n\nI'm not sure what kind of benefit you might expect to obtain from just a slope prediction? A slope target would necessitate an intercept target as well, and you would still need to assign an uncertainty to it...so that's 3 values to predict, which is more than the 2 already needed. Going completely linear may also mean you're missing out on a more complex function if you manage to find good features...",
    "958917": "In CV it seems to me that slope is easier to predict that the value, but the percent variation is even better"
  },
  "source": "meta"
}