{ "id": 228796, "title": "Estimate of best prediction score ", "url": "/competitions/indoor-location-navigation/discussion/228796", "author_name": "Kamal Das", "post_date": "2021-03-26T14:52:19.291000", "votes": 13, "comment_count": 15, "views": 0, "content": "

Hi All,

\n

This has been a crazy competition for me. I was looking at my submissions and realised a month back I had ticked my scores around 8 as the best submissions think there was no way I would get better 😄

\n

Everyone has helped push us to do better, and the best public NB has a score of 4.762!! And I am so surprised I was able to be in 4. something score!!

\n

Some suggestions from everyone : @nooblife 's on CV strategy, (https://www.kaggle.com/c/indoor-location-navigation/discussion/228199)
\n@seuguh on site-specific model will do better ( https://www.kaggle.com/c/indoor-location-navigation/discussion/220835 ); notebooks on using the locations to map and adjust the locations slightly .. (https://www.kaggle.com/c/indoor-location-navigation/discussion/221166) many many helped…

\n

and I am sure the top score will keep improving from its current score of 3.621 by @vaghefi

\n

A quick side game, if you will, what do you think will be the best prediction score around 17th of May when the submission will end? How far do you think we may do?

\n

I think the gold medal teams may be in the 2.25 to 2.5 band; and we will see a lot of improvements!

\n

What do you think??… put in your predictions and let us see how they turn up in another 6 odd weeks!

", "messages": [ { "id": 1253279, "postDate": "2021-03-26T14:52:19.293Z", "content": "

Hi All,

\n

This has been a crazy competition for me. I was looking at my submissions and realised a month back I had ticked my scores around 8 as the best submissions think there was no way I would get better 😄

\n

Everyone has helped push us to do better, and the best public NB has a score of 4.762!! And I am so surprised I was able to be in 4. something score!!

\n

Some suggestions from everyone : @nooblife 's on CV strategy, (https://www.kaggle.com/c/indoor-location-navigation/discussion/228199)
\n@seuguh on site-specific model will do better ( https://www.kaggle.com/c/indoor-location-navigation/discussion/220835 ); notebooks on using the locations to map and adjust the locations slightly .. (https://www.kaggle.com/c/indoor-location-navigation/discussion/221166) many many helped…

\n

and I am sure the top score will keep improving from its current score of 3.621 by @vaghefi

\n

A quick side game, if you will, what do you think will be the best prediction score around 17th of May when the submission will end? How far do you think we may do?

\n

I think the gold medal teams may be in the 2.25 to 2.5 band; and we will see a lot of improvements!

\n

What do you think??… put in your predictions and let us see how they turn up in another 6 odd weeks!

", "rawMarkdown": "Hi All,\n\nThis has been a crazy competition for me. I was looking at my submissions and realised a month back I had ticked my scores around 8 as the best submissions think there was no way I would get better 😄\n\nEveryone has helped push us to do better, and the best public NB has a score of 4.762!! And I am so surprised I was able to be in 4. something score!!\n\nSome suggestions from everyone : @nooblife 's on CV strategy, (https://www.kaggle.com/c/indoor-location-navigation/discussion/228199)\n@seuguh on site-specific model will do better ( https://www.kaggle.com/c/indoor-location-navigation/discussion/220835 ); notebooks on using the locations to map and adjust the locations slightly .. (https://www.kaggle.com/c/indoor-location-navigation/discussion/221166) many many helped... \n\nand I am sure the top score will keep improving from its current score of 3.621 by @vaghefi\n\nA quick side game, if you will, what do you think will be the best prediction score around 17th of May when the submission will end? How far do you think we may do?\n\nI think the gold medal teams may be in the 2.25 to 2.5 band; and we will see a lot of improvements!\n\nWhat do you think??... put in your predictions and let us see how they turn up in another 6 odd weeks!\n", "votes": 13 }, { "id": 1253329, "postDate": "2021-03-26T16:11:13.703Z", "content": "

I remember telling my self, one month ago, that scores would never go below 5. It's surprising how we can squeeze data here!
\nMy bet is for a gold zone of 2.6-2.9 (public LB scores).

", "rawMarkdown": "I remember telling my self, one month ago, that scores would never go below 5. It's surprising how we can squeeze data here!\nMy bet is for a gold zone of 2.6-2.9 (public LB scores).", "votes": 3, "replies": [ { "id": 1253332, "postDate": "2021-03-26T16:13:13.223Z", "content": "

Look forward to your sub 3 scores @zidmie 😄

", "rawMarkdown": "Look forward to your sub 3 scores @zidmie 😄", "votes": 2 }, { "id": 1257635, "postDate": "2021-03-31T02:24:03.723Z", "content": "

Currently, most are relying on wifi RSSI, corrected by the sensors.
\nI suspect when the submissions are the other way around (rely on accelerometer and gyroscope corrected by wifi) then I suspect it should be accurate to 1.5 - 2 strides which I believe is between 1 and 2 by the error metric.

", "rawMarkdown": "Currently, most are relying on wifi RSSI, corrected by the sensors.\nI suspect when the submissions are the other way around (rely on accelerometer and gyroscope corrected by wifi) then I suspect it should be accurate to 1.5 - 2 strides which I believe is between 1 and 2 by the error metric.", "votes": 1 }, { "id": 1258093, "postDate": "2021-03-31T10:42:33.533Z", "content": "

if you've got the correct floor, then indeed the error metric corresponds to Euclidean distance in metres between the predicted and true positions.

", "rawMarkdown": "if you've got the correct floor, then indeed the error metric corresponds to Euclidean distance in metres between the predicted and true positions.", "votes": 1 }, { "id": 1259584, "postDate": "2021-04-01T14:19:29.420Z", "content": "

I will try my best, but 1.5 - 2 looks very hard 😭

", "rawMarkdown": "I will try my best, but 1.5 - 2 looks very hard 😭", "votes": 3 }, { "id": 1259598, "postDate": "2021-04-01T14:30:34.557Z", "content": "

😄 We are all looking at you @mamasinkgs !!

\n

Also, what is your magic trick? what are you doing that all of us are missing … you are leagues ahead of all of us! All the best on your expected sub 3 score soon!! 👍🙌

", "rawMarkdown": "😄 We are all looking at you @mamasinkgs !!\n\nAlso, what is your magic trick? what are you doing that all of us are missing ... you are leagues ahead of all of us! All the best on your expected sub 3 score soon!! 👍🙌", "votes": 4 }, { "id": 1259642, "postDate": "2021-04-01T14:58:59.147Z", "content": "

Thank you. Of course I can't say what the magic is, but I think these 2 steps are very important in this competition.
\n(1) EDA by yourself.
\n(2) Understand good kernels (regardless of their score).

", "rawMarkdown": "Thank you. Of course I can't say what the magic is, but I think these 2 steps are very important in this competition.\n(1) EDA by yourself. \n(2) Understand good kernels (regardless of their score).\n", "votes": 5 }, { "id": 1262836, "postDate": "2021-04-04T19:05:59.717Z", "content": "

@jbomitchell yes, in terms of meters, but I have not calculated the average stride length

", "rawMarkdown": "@jbomitchell yes, in terms of meters, but I have not calculated the average stride length", "votes": 1 }, { "id": 1262837, "postDate": "2021-04-04T19:11:45.663Z", "content": "

@nigelhenry - probably around 75 cm.

", "rawMarkdown": "@nigelhenry - probably around 75 cm.", "votes": 1 } ] }, { "id": 1279421, "postDate": "2021-04-20T22:29:03.107Z", "content": "

A possible caveat is that these models have been selected and evolved on the basis of LB feedback from their scores on the public LB. It's possible that performance on the private LB doesn't always follow that, and that some good-scoring models may turn out to be overfitted to the public portion of the test data. As a specific example, there seems to be a broad consensus over the floor predictions for the public LB paths, but I'm aware of a number of uncertainties when it comes to the floor predictions for the private LB.

", "rawMarkdown": "A possible caveat is that these models have been selected and evolved on the basis of LB feedback from their scores on the public LB. It's possible that performance on the private LB doesn't always follow that, and that some good-scoring models may turn out to be overfitted to the public portion of the test data. As a specific example, there seems to be a broad consensus over the floor predictions for the public LB paths, but I'm aware of a number of uncertainties when it comes to the floor predictions for the private LB.", "votes": 1, "replies": [ { "id": 1279572, "postDate": "2021-04-21T04:21:20.953Z", "content": "

agree @jbomitchell … don't think many are changing the floor in post process… and public LB is only on 15% of the data….

\n

makes me wonder….
\nimage

\n

Link if image does not show: https://www.kaggle.com/kmldas/gif-for-discussion?select=lb.JPG

", "rawMarkdown": "agree @jbomitchell ... don't think many are changing the floor in post process... and public LB is only on 15% of the data....\n\n\nmakes me wonder....\n[image](https://storage.googleapis.com/kagglesdsdata/datasets/1277000/2128116/lb.JPG?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=databundle-worker-v2%40kaggle-161607.iam.gserviceaccount.com%2F20210419%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20210419T162539Z&X-Goog-Expires=172799&X-Goog-SignedHeaders=host&X-Goog-Signature=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)\n\nLink if image does not show: https://www.kaggle.com/kmldas/gif-for-discussion?select=lb.JPG\n\n", "votes": 1 } ] }, { "id": 1253345, "postDate": "2021-03-26T16:23:39.510Z", "content": "

What strikes me is the extent to which the Kaggle community has effectively solved this problem - we are now down to locating a phone in a multi-storey building to within an accuracy of about 3.65 metres. Well done to all of you folk who've contributed to this!

", "rawMarkdown": "What strikes me is the extent to which the Kaggle community has effectively solved this problem - we are now down to locating a phone in a multi-storey building to within an accuracy of about 3.65 metres. Well done to all of you folk who've contributed to this!", "votes": 2, "replies": [ { "id": 1253373, "postDate": "2021-03-26T16:50:55.707Z", "content": "

Agree! Was totally not expecting it… proves how well the Kaggle community can solve seemingly complicated problems and get such great levels of accuracy!!

", "rawMarkdown": "Agree! Was totally not expecting it... proves how well the Kaggle community can solve seemingly complicated problems and get such great levels of accuracy!! ", "votes": 1 } ] }, { "id": 1278238, "postDate": "2021-04-19T17:13:17.040Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 1284042, "postDate": "2021-04-25T14:14:35.377Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 1253329, "author_name": "Zidmie", "author_url": "", "post_date": "2021-03-26T16:11:13.703000", "content": "

I remember telling my self, one month ago, that scores would never go below 5. It's surprising how we can squeeze data here!
\nMy bet is for a gold zone of 2.6-2.9 (public LB scores).

", "votes": 3, "replies": [ { "id": 1253332, "author_name": "Kamal Das", "author_url": "", "post_date": "2021-03-26T16:13:13.223000", "content": "

Look forward to your sub 3 scores @zidmie 😄

", "votes": 2, "replies": [] }, { "id": 1257635, "author_name": "Nigel A. R. Henry", "author_url": "", "post_date": "2021-03-31T02:24:03.723000", "content": "

Currently, most are relying on wifi RSSI, corrected by the sensors.
\nI suspect when the submissions are the other way around (rely on accelerometer and gyroscope corrected by wifi) then I suspect it should be accurate to 1.5 - 2 strides which I believe is between 1 and 2 by the error metric.

", "votes": 1, "replies": [] }, { "id": 1258093, "author_name": "John Mitchell", "author_url": "", "post_date": "2021-03-31T10:42:33.533000", "content": "

if you've got the correct floor, then indeed the error metric corresponds to Euclidean distance in metres between the predicted and true positions.

", "votes": 1, "replies": [] }, { "id": 1259584, "author_name": "mamas", "author_url": "", "post_date": "2021-04-01T14:19:29.420000", "content": "

I will try my best, but 1.5 - 2 looks very hard 😭

", "votes": 3, "replies": [] }, { "id": 1259598, "author_name": "Kamal Das", "author_url": "", "post_date": "2021-04-01T14:30:34.557000", "content": "

😄 We are all looking at you @mamasinkgs !!

\n

Also, what is your magic trick? what are you doing that all of us are missing … you are leagues ahead of all of us! All the best on your expected sub 3 score soon!! 👍🙌

", "votes": 4, "replies": [] }, { "id": 1259642, "author_name": "mamas", "author_url": "", "post_date": "2021-04-01T14:58:59.147000", "content": "

Thank you. Of course I can't say what the magic is, but I think these 2 steps are very important in this competition.
\n(1) EDA by yourself.
\n(2) Understand good kernels (regardless of their score).

", "votes": 5, "replies": [] }, { "id": 1262836, "author_name": "Nigel A. R. Henry", "author_url": "", "post_date": "2021-04-04T19:05:59.717000", "content": "

@jbomitchell yes, in terms of meters, but I have not calculated the average stride length

", "votes": 1, "replies": [] }, { "id": 1262837, "author_name": "John Mitchell", "author_url": "", "post_date": "2021-04-04T19:11:45.663000", "content": "

@nigelhenry - probably around 75 cm.

", "votes": 1, "replies": [] } ] }, { "id": 1279421, "author_name": "John Mitchell", "author_url": "", "post_date": "2021-04-20T22:29:03.107000", "content": "

A possible caveat is that these models have been selected and evolved on the basis of LB feedback from their scores on the public LB. It's possible that performance on the private LB doesn't always follow that, and that some good-scoring models may turn out to be overfitted to the public portion of the test data. As a specific example, there seems to be a broad consensus over the floor predictions for the public LB paths, but I'm aware of a number of uncertainties when it comes to the floor predictions for the private LB.

", "votes": 1, "replies": [ { "id": 1279572, "author_name": "Kamal Das", "author_url": "", "post_date": "2021-04-21T04:21:20.953000", "content": "

agree @jbomitchell … don't think many are changing the floor in post process… and public LB is only on 15% of the data….

\n

makes me wonder….
\nimage

\n

Link if image does not show: https://www.kaggle.com/kmldas/gif-for-discussion?select=lb.JPG

", "votes": 1, "replies": [] } ] }, { "id": 1253345, "author_name": "John Mitchell", "author_url": "", "post_date": "2021-03-26T16:23:39.510000", "content": "

What strikes me is the extent to which the Kaggle community has effectively solved this problem - we are now down to locating a phone in a multi-storey building to within an accuracy of about 3.65 metres. Well done to all of you folk who've contributed to this!

", "votes": 2, "replies": [ { "id": 1253373, "author_name": "Kamal Das", "author_url": "", "post_date": "2021-03-26T16:50:55.707000", "content": "

Agree! Was totally not expecting it… proves how well the Kaggle community can solve seemingly complicated problems and get such great levels of accuracy!!

", "votes": 1, "replies": [] } ] }, { "id": 1278238, "author_name": "", "author_url": "", "post_date": "2021-04-19T17:13:17.040000", "content": "", "votes": 0, "replies": [ { "id": 1284042, "author_name": "", "author_url": "", "post_date": "2021-04-25T14:14:35.377000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1253279": "Hi All,\n\nThis has been a crazy competition for me. I was looking at my submissions and realised a month back I had ticked my scores around 8 as the best submissions think there was no way I would get better 😄\n\nEveryone has helped push us to do better, and the best public NB has a score of 4.762!! And I am so surprised I was able to be in 4. something score!!\n\nSome suggestions from everyone : @nooblife 's on CV strategy, (https://www.kaggle.com/c/indoor-location-navigation/discussion/228199)\n@seuguh on site-specific model will do better ( https://www.kaggle.com/c/indoor-location-navigation/discussion/220835 ); notebooks on using the locations to map and adjust the locations slightly .. (https://www.kaggle.com/c/indoor-location-navigation/discussion/221166) many many helped... \n\nand I am sure the top score will keep improving from its current score of 3.621 by @vaghefi\n\nA quick side game, if you will, what do you think will be the best prediction score around 17th of May when the submission will end? How far do you think we may do?\n\nI think the gold medal teams may be in the 2.25 to 2.5 band; and we will see a lot of improvements!\n\nWhat do you think??... put in your predictions and let us see how they turn up in another 6 odd weeks!\n", "1253329": "I remember telling my self, one month ago, that scores would never go below 5. It's surprising how we can squeeze data here!\nMy bet is for a gold zone of 2.6-2.9 (public LB scores).", "1279421": "A possible caveat is that these models have been selected and evolved on the basis of LB feedback from their scores on the public LB. It's possible that performance on the private LB doesn't always follow that, and that some good-scoring models may turn out to be overfitted to the public portion of the test data. As a specific example, there seems to be a broad consensus over the floor predictions for the public LB paths, but I'm aware of a number of uncertainties when it comes to the floor predictions for the private LB.", "1253345": "What strikes me is the extent to which the Kaggle community has effectively solved this problem - we are now down to locating a phone in a multi-storey building to within an accuracy of about 3.65 metres. Well done to all of you folk who've contributed to this!", "1278238": "" } }