{ "id": 237542, "title": "Do \"you\" expect a Shakeup?", "url": "/competitions/indoor-location-navigation/discussion/237542", "author_name": "", "post_date": "2021-05-09T09:17:31.659663800Z", "votes": 10, "comment_count": 7, "views": 0, "content": "

Hi,

\n

What are your expectations? Do \"you\" expect a Shakeup?
\n@mamasinkgs suggests some are hiding true scores by using different floors etc
\nonly 15% of data is used to calculate the Public LB

\n

what are your thoughts?

\n

Anyway,,, all the best in the final week of the competition..
\nit's been a hectic comp and can;t wait for it to get over and get some rest!!!

", "messages": [ { "id": "1298852", "postDate": "05/09/2021 09:17:31", "content": "

Hi,

\n

What are your expectations? Do \"you\" expect a Shakeup?
\n@mamasinkgs suggests some are hiding true scores by using different floors etc
\nonly 15% of data is used to calculate the Public LB

\n

what are your thoughts?

\n

Anyway,,, all the best in the final week of the competition..
\nit's been a hectic comp and can;t wait for it to get over and get some rest!!!

", "rawMarkdown": "Hi,\n\nWhat are your expectations? Do \"you\" expect a Shakeup?\n@mamasinkgs suggests some are hiding true scores by using different floors etc\nonly 15% of data is used to calculate the Public LB\n\nwhat are your thoughts?\n\nAnyway,,, all the best in the final week of the competition.. \nit's been a hectic comp and can;t wait for it to get over and get some rest!!!", "votes": null }, { "id": "1299045", "postDate": "05/09/2021 12:37:59", "content": "

I think shakeup won't be so big for the teams whose floor predictions are correct. However, If your CV-LB is not correlated and you made many submissions (e.g. 300 submissons), you will get much worse score in private LB. If your floor predictions are correct and your CV-LB is correlated, you will be happy.

", "rawMarkdown": "I think shakeup won't be so big for the teams whose floor predictions are correct. However, If your CV-LB is not correlated and you made many submissions (e.g. 300 submissons), you will get much worse score in private LB. If your floor predictions are correct and your CV-LB is correlated, you will be happy.", "votes": null }, { "id": "1299128", "postDate": "05/09/2021 13:45:49", "content": "

Also, looking forward to @zidmie and you becoming GMs once the LB is finalised!! 😄
\ni am sure you both have solid CVs
\nMy CV is more shaky… and fingers crossed i don;t fall off too much 😊

", "rawMarkdown": "Also, looking forward to @zidmie and you becoming GMs once the LB is finalised!! 😄 \ni am sure you both have solid CVs\nMy CV is more shaky... and fingers crossed i don;t fall off too much 😊", "votes": null }, { "id": "1299515", "postDate": "05/09/2021 18:19:32", "content": "
\n

Hi,

\n

What are your expectations? Do \"you\" expect a Shakeup?

\n
\n

Yes, I expect a shakeup. I expect that many of us have unintentionally overfitted to the public LB.

\n
\n

@mamasinkgs suggests some are hiding true scores by using different floors etc
\n only 15% of data is used to calculate the Public LB

\n

what are your thoughts?

\n
\n

Perhaps one or two, but not many. I don't see much point in sandbagging in a competition like this.

\n
\n

Anyway,,, all the best in the final week of the competition..
\n it's been a hectic comp and can;t wait for it to get over and get some rest!!!

\n
\n

It would be interesting to think about how people might plan to utilise the availability of two counting submissions.

", "rawMarkdown": "> Hi,\n> \n> What are your expectations? Do \"you\" expect a Shakeup?\n\nYes, I expect a shakeup. I expect that many of us have unintentionally overfitted to the public LB.\n\n> @mamasinkgs suggests some are hiding true scores by using different floors etc\n> only 15% of data is used to calculate the Public LB\n> \n> what are your thoughts?\n\nPerhaps one or two, but not many. I don't see much point in sandbagging in a competition like this.\n\n> \n> Anyway,,, all the best in the final week of the competition.. \n> it's been a hectic comp and can;t wait for it to get over and get some rest!!!\n\nIt would be interesting to think about how people might plan to utilise the availability of *two* counting submissions.", "votes": null }, { "id": "1299520", "postDate": "05/09/2021 18:23:10", "content": "

yes .. choosing the final two will be interesting-- what is your plan for this?

\n

i currently plan to use two different floor plans! … lets see.. one more week to go before i decide on this!

", "rawMarkdown": "yes .. choosing the final two will be interesting-- what is your plan for this?\n\ni currently plan to use two different floor plans! ... lets see.. one more week to go before i decide on this!", "votes": null }, { "id": "1299525", "postDate": "05/09/2021 18:28:00", "content": "

Yes, overfitting to the public LB is definitely easy in this competition - I've had trouble doing proper CV because the test and train data are all so intertwined, so it's difficult to get a real validation split - so we'll see! :)

", "rawMarkdown": "Yes, overfitting to the public LB is definitely easy in this competition - I've had trouble doing proper CV because the test and train data are all so intertwined, so it's difficult to get a real validation split - so we'll see! :)", "votes": null }, { "id": "1299531", "postDate": "05/09/2021 18:35:15", "content": "

Agree, proper CV is difficult, especially because not many (if any) paths overlap.

", "rawMarkdown": "Agree, proper CV is difficult, especially because not many (if any) paths overlap.", "votes": null }, { "id": "1299636", "postDate": "05/09/2021 22:16:50", "content": "

Two different floor plans seems like a decent call. It might mean having two models with distinct choices of floors burned into them, and at the very least snapping to separate sets of grid points for each one. Or the lazy approach would be to paste different floors into the same coordinates.

\n

There are obviously other ways to use the second submission. I can immediately think of a couple - one of which is definitely a possibility for me to pursue (but it currently scores too poorly to submit). People on here are doubtless smart enough to think up and assess all sorts of protocols for leveraging the second submission.

", "rawMarkdown": "Two different floor plans seems like a decent call. It might mean having two models with distinct choices of floors burned into them, and at the very least snapping to separate sets of grid points for each one. Or the lazy approach would be to paste different floors into the same coordinates. \n\nThere are obviously other ways to use the second submission. I can immediately think of a couple - one of which is definitely a possibility for me to pursue (but it currently scores too poorly to submit). People on here are doubtless smart enough to think up and assess all sorts of protocols for leveraging the second submission.", "votes": null } ], "comments": [ { "id": 1299045, "author_name": "mamasinkgs", "author_url": "", "post_date": "05/09/2021 12:37:59", "content": "

I think shakeup won't be so big for the teams whose floor predictions are correct. However, If your CV-LB is not correlated and you made many submissions (e.g. 300 submissons), you will get much worse score in private LB. If your floor predictions are correct and your CV-LB is correlated, you will be happy.

", "votes": null, "replies": [ { "id": 1299128, "author_name": "kmldas", "author_url": "", "post_date": "05/09/2021 13:45:49", "content": "

Also, looking forward to @zidmie and you becoming GMs once the LB is finalised!! 😄
\ni am sure you both have solid CVs
\nMy CV is more shaky… and fingers crossed i don;t fall off too much 😊

", "votes": null, "replies": [] } ] }, { "id": 1299515, "author_name": "jbomitchell", "author_url": "", "post_date": "05/09/2021 18:19:32", "content": "
\n

Hi,

\n

What are your expectations? Do \"you\" expect a Shakeup?

\n
\n

Yes, I expect a shakeup. I expect that many of us have unintentionally overfitted to the public LB.

\n
\n

@mamasinkgs suggests some are hiding true scores by using different floors etc
\n only 15% of data is used to calculate the Public LB

\n

what are your thoughts?

\n
\n

Perhaps one or two, but not many. I don't see much point in sandbagging in a competition like this.

\n
\n

Anyway,,, all the best in the final week of the competition..
\n it's been a hectic comp and can;t wait for it to get over and get some rest!!!

\n
\n

It would be interesting to think about how people might plan to utilise the availability of two counting submissions.

", "votes": null, "replies": [ { "id": 1299520, "author_name": "kmldas", "author_url": "", "post_date": "05/09/2021 18:23:10", "content": "

yes .. choosing the final two will be interesting-- what is your plan for this?

\n

i currently plan to use two different floor plans! … lets see.. one more week to go before i decide on this!

", "votes": null, "replies": [] }, { "id": 1299525, "author_name": "chris62", "author_url": "", "post_date": "05/09/2021 18:28:00", "content": "

Yes, overfitting to the public LB is definitely easy in this competition - I've had trouble doing proper CV because the test and train data are all so intertwined, so it's difficult to get a real validation split - so we'll see! :)

", "votes": null, "replies": [] }, { "id": 1299531, "author_name": "olaf2000", "author_url": "", "post_date": "05/09/2021 18:35:15", "content": "

Agree, proper CV is difficult, especially because not many (if any) paths overlap.

", "votes": null, "replies": [] }, { "id": 1299636, "author_name": "jbomitchell", "author_url": "", "post_date": "05/09/2021 22:16:50", "content": "

Two different floor plans seems like a decent call. It might mean having two models with distinct choices of floors burned into them, and at the very least snapping to separate sets of grid points for each one. Or the lazy approach would be to paste different floors into the same coordinates.

\n

There are obviously other ways to use the second submission. I can immediately think of a couple - one of which is definitely a possibility for me to pursue (but it currently scores too poorly to submit). People on here are doubtless smart enough to think up and assess all sorts of protocols for leveraging the second submission.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1298852": "Hi,\n\nWhat are your expectations? Do \"you\" expect a Shakeup?\n@mamasinkgs suggests some are hiding true scores by using different floors etc\nonly 15% of data is used to calculate the Public LB\n\nwhat are your thoughts?\n\nAnyway,,, all the best in the final week of the competition.. \nit's been a hectic comp and can;t wait for it to get over and get some rest!!!", "1299045": "I think shakeup won't be so big for the teams whose floor predictions are correct. However, If your CV-LB is not correlated and you made many submissions (e.g. 300 submissons), you will get much worse score in private LB. If your floor predictions are correct and your CV-LB is correlated, you will be happy.", "1299128": "Also, looking forward to @zidmie and you becoming GMs once the LB is finalised!! 😄 \ni am sure you both have solid CVs\nMy CV is more shaky... and fingers crossed i don;t fall off too much 😊", "1299515": "> Hi,\n> \n> What are your expectations? Do \"you\" expect a Shakeup?\n\nYes, I expect a shakeup. I expect that many of us have unintentionally overfitted to the public LB.\n\n> @mamasinkgs suggests some are hiding true scores by using different floors etc\n> only 15% of data is used to calculate the Public LB\n> \n> what are your thoughts?\n\nPerhaps one or two, but not many. I don't see much point in sandbagging in a competition like this.\n\n> \n> Anyway,,, all the best in the final week of the competition.. \n> it's been a hectic comp and can;t wait for it to get over and get some rest!!!\n\nIt would be interesting to think about how people might plan to utilise the availability of *two* counting submissions.", "1299520": "yes .. choosing the final two will be interesting-- what is your plan for this?\n\ni currently plan to use two different floor plans! ... lets see.. one more week to go before i decide on this!", "1299525": "Yes, overfitting to the public LB is definitely easy in this competition - I've had trouble doing proper CV because the test and train data are all so intertwined, so it's difficult to get a real validation split - so we'll see! :)", "1299531": "Agree, proper CV is difficult, especially because not many (if any) paths overlap.", "1299636": "Two different floor plans seems like a decent call. It might mean having two models with distinct choices of floors burned into them, and at the very least snapping to separate sets of grid points for each one. Or the lazy approach would be to paste different floors into the same coordinates. \n\nThere are obviously other ways to use the second submission. I can immediately think of a couple - one of which is definitely a possibility for me to pursue (but it currently scores too poorly to submit). People on here are doubtless smart enough to think up and assess all sorts of protocols for leveraging the second submission." }, "source": "meta" }