I've run a couple test submissions. It seems RMSLE on LB is consistently better if you submit a float prediction instead of an integer.
", "messages": [ { "id": "130399", "postDate": "08/06/2016 16:37:37", "content": "I've run a couple test submissions. It seems RMSLE on LB is consistently better if you submit a float prediction instead of an integer.
", "rawMarkdown": "I've run a couple test submissions. It seems RMSLE on LB is consistently better if you submit a float prediction instead of an integer.", "votes": null }, { "id": "130422", "postDate": "08/06/2016 21:33:03", "content": "Is your integer prediction a 'floor' or 'ceiling' version of your float prediction? If you are submitting the floor version, try submitting the ceiling version. I suspect it will then give you the same result.
", "rawMarkdown": "Is your integer prediction a 'floor' or 'ceiling' version of your float prediction? If you are submitting the floor version, try submitting the ceiling version. I suspect it will then give you the same result.", "votes": null }, { "id": "130427", "postDate": "08/06/2016 22:40:04", "content": "Thanks ... will give ceiling a try.
\n\nI also built a special rounding function checked for the nearest integer on a log(p+1) basis. Basically, took a floor and ceiling of my prediction and used the one that had lower log error from my float prediction. This also made the score worse, rather than better.
", "rawMarkdown": "Thanks ... will give ceiling a try. \r\n\r\nI also built a special rounding function checked for the nearest integer on a log(p+1) basis. Basically, took a floor and ceiling of my prediction and used the one that had lower log error from my float prediction. This also made the score worse, rather than better.", "votes": null }, { "id": "130564", "postDate": "08/08/2016 15:06:02", "content": "@crimeprophet : Just tried it, ceiling is also significantly worse than submitting floats. So, unless someone has seen or heard otherwise, you are best off submitting with a few decimal places.
", "rawMarkdown": "crimeprophet : Just tried it, ceiling is also significantly worse than submitting floats. So, unless someone has seen or heard otherwise, you are best off submitting with a few decimal places.", "votes": null }, { "id": "130575", "postDate": "08/08/2016 17:03:57", "content": "Most/all Kaggle numerical competitions expect floating point numbers - in this case they are processed by log(x + 1) anyway and will become floating point.
", "rawMarkdown": "Most/all Kaggle numerical competitions expect floating point numbers - in this case they are processed by log(x + 1) anyway and will become floating point.", "votes": null }, { "id": "130578", "postDate": "08/08/2016 17:18:40", "content": "@happycube - Yes, but instructions made some implication we should submit integers. And there was confusion on the point in the "Welcome" thread.
", "rawMarkdown": "happycube - Yes, but instructions made some implication we should submit integers. And there was confusion on the point in the \"Welcome\" thread.", "votes": null } ], "comments": [ { "id": 130422, "author_name": "crimeprophet", "author_url": "", "post_date": "08/06/2016 21:33:03", "content": "Is your integer prediction a 'floor' or 'ceiling' version of your float prediction? If you are submitting the floor version, try submitting the ceiling version. I suspect it will then give you the same result.
", "votes": null, "replies": [] }, { "id": 130427, "author_name": "patcoady", "author_url": "", "post_date": "08/06/2016 22:40:04", "content": "Thanks ... will give ceiling a try.
\n\nI also built a special rounding function checked for the nearest integer on a log(p+1) basis. Basically, took a floor and ceiling of my prediction and used the one that had lower log error from my float prediction. This also made the score worse, rather than better.
", "votes": null, "replies": [] }, { "id": 130564, "author_name": "patcoady", "author_url": "", "post_date": "08/08/2016 15:06:02", "content": "@crimeprophet : Just tried it, ceiling is also significantly worse than submitting floats. So, unless someone has seen or heard otherwise, you are best off submitting with a few decimal places.
", "votes": null, "replies": [] }, { "id": 130575, "author_name": "happycube", "author_url": "", "post_date": "08/08/2016 17:03:57", "content": "Most/all Kaggle numerical competitions expect floating point numbers - in this case they are processed by log(x + 1) anyway and will become floating point.
", "votes": null, "replies": [] }, { "id": 130578, "author_name": "patcoady", "author_url": "", "post_date": "08/08/2016 17:18:40", "content": "@happycube - Yes, but instructions made some implication we should submit integers. And there was confusion on the point in the "Welcome" thread.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "130399": "I've run a couple test submissions. It seems RMSLE on LB is consistently better if you submit a float prediction instead of an integer.", "130422": "Is your integer prediction a 'floor' or 'ceiling' version of your float prediction? If you are submitting the floor version, try submitting the ceiling version. I suspect it will then give you the same result.", "130427": "Thanks ... will give ceiling a try. \r\n\r\nI also built a special rounding function checked for the nearest integer on a log(p+1) basis. Basically, took a floor and ceiling of my prediction and used the one that had lower log error from my float prediction. This also made the score worse, rather than better.", "130564": "crimeprophet : Just tried it, ceiling is also significantly worse than submitting floats. So, unless someone has seen or heard otherwise, you are best off submitting with a few decimal places.", "130575": "Most/all Kaggle numerical competitions expect floating point numbers - in this case they are processed by log(x + 1) anyway and will become floating point.", "130578": "happycube - Yes, but instructions made some implication we should submit integers. And there was confusion on the point in the \"Welcome\" thread." }, "source": "meta" }