I am using a direct method to estimate the l.v volume from each folders, without incorporating any machine learning techniques. Hence, instead of getting a predicted continuous distribution in the form of a sigmoid curve , I am getting a particular volume corresponding to the l.v of each patient in the form of a step function, and then calculating the CRPS based on the given expression. Is there any requirement for the contest to necessarily compute the l.v volume in terms of the continuous distribution function?
", "messages": [ { "id": "108255", "postDate": "02/16/2016 13:53:59", "content": "I am using a direct method to estimate the l.v volume from each folders, without incorporating any machine learning techniques. Hence, instead of getting a predicted continuous distribution in the form of a sigmoid curve , I am getting a particular volume corresponding to the l.v of each patient in the form of a step function, and then calculating the CRPS based on the given expression. Is there any requirement for the contest to necessarily compute the l.v volume in terms of the continuous distribution function?
", "rawMarkdown": "I am using a direct method to estimate the l.v volume from each folders, without incorporating any machine learning techniques. Hence, instead of getting a predicted continuous distribution in the form of a sigmoid curve , I am getting a particular volume corresponding to the l.v of each patient in the form of a step function, and then calculating the CRPS based on the given expression. Is there any requirement for the contest to necessarily compute the l.v volume in terms of the continuous distribution function?", "votes": null }, { "id": "108264", "postDate": "02/16/2016 14:27:24", "content": "My understanding is that yes your "continuous distribution function" can indeed be a step function if you wish. However, odds are one would best your score by constructing sigmoids from the exact same predictions.
", "rawMarkdown": "My understanding is that yes your \"continuous distribution function\" can indeed be a step function if you wish. However, odds are one would best your score by constructing sigmoids from the exact same predictions.", "votes": null }, { "id": "108294", "postDate": "02/16/2016 16:08:47", "content": "@Woolsey , could you please tell what do you exactly mean by "constructing sigmoid" from step functions? How can I just fit any sigmoid with arbitrary parameters on a step function?
", "rawMarkdown": "Woolsey , could you please tell what do you exactly mean by \"**constructing sigmoid**\" from step functions? How can I just fit any sigmoid with arbitrary parameters on a step function?", "votes": null }, { "id": "108303", "postDate": "02/16/2016 16:38:09", "content": "There are several possibilities. I would suggest using y=1/(1+e^c.(x-v)), where v is the volume you predicted and c a constant which describe how "step" the sigmoid is (here it must be negative, likely above -5). To fine-tune c to the level of noise you have, you can either run a simulation or test with the leaderboard.
", "rawMarkdown": "There are several possibilities. I would suggest using y=1/(1+e^c.(x-v)), where v is the volume you predicted and c a constant which describe how \"step\" the sigmoid is (here it must be negative, likely above -5). To fine-tune c to the level of noise you have, you can either run a simulation or test with the leaderboard.", "votes": null } ], "comments": [ { "id": 108264, "author_name": "woolsey", "author_url": "", "post_date": "02/16/2016 14:27:24", "content": "My understanding is that yes your "continuous distribution function" can indeed be a step function if you wish. However, odds are one would best your score by constructing sigmoids from the exact same predictions.
", "votes": null, "replies": [] }, { "id": 108294, "author_name": "", "author_url": "", "post_date": "02/16/2016 16:08:47", "content": "@Woolsey , could you please tell what do you exactly mean by "constructing sigmoid" from step functions? How can I just fit any sigmoid with arbitrary parameters on a step function?
", "votes": null, "replies": [] }, { "id": 108303, "author_name": "woolsey", "author_url": "", "post_date": "02/16/2016 16:38:09", "content": "There are several possibilities. I would suggest using y=1/(1+e^c.(x-v)), where v is the volume you predicted and c a constant which describe how "step" the sigmoid is (here it must be negative, likely above -5). To fine-tune c to the level of noise you have, you can either run a simulation or test with the leaderboard.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "108255": "I am using a direct method to estimate the l.v volume from each folders, without incorporating any machine learning techniques. Hence, instead of getting a predicted continuous distribution in the form of a sigmoid curve , I am getting a particular volume corresponding to the l.v of each patient in the form of a step function, and then calculating the CRPS based on the given expression. Is there any requirement for the contest to necessarily compute the l.v volume in terms of the continuous distribution function?", "108264": "My understanding is that yes your \"continuous distribution function\" can indeed be a step function if you wish. However, odds are one would best your score by constructing sigmoids from the exact same predictions.", "108294": "Woolsey , could you please tell what do you exactly mean by \"**constructing sigmoid**\" from step functions? How can I just fit any sigmoid with arbitrary parameters on a step function?", "108303": "There are several possibilities. I would suggest using y=1/(1+e^c.(x-v)), where v is the volume you predicted and c a constant which describe how \"step\" the sigmoid is (here it must be negative, likely above -5). To fine-tune c to the level of noise you have, you can either run a simulation or test with the leaderboard." }, "source": "meta" }