{ "id": 263733, "title": "Smartphone competition shakeup scatterplot", "url": "/competitions/google-smartphone-decimeter-challenge/discussion/263733", "author_name": "", "post_date": "2021-08-10T05:41:43.429686500Z", "votes": 8, "comment_count": 2, "views": 0, "content": "

Here is the shakeup for the Google Smartphone Decimeter Challenge (produced using the Shakeup interactive scatterplot maker. Right click to see an enlarged version, or see the scatterplot maker to play with an interactive version)

\n

\"\"

\n

As suggested by @t88take in the topic \"Shake up or not shake up?\" things were not too bad in the end. There were a few notable exceptions; there is a cloud of shakedowns whose public score was in the range 40-120, that dropped down to the 300-400 range. I suspect these were submissions of over-fitted ensembles. There are a good number of isoscore strings formed by people all submitting exactly the same public notebook, for example 31 teams went with the [GSDC] Smart Ensembling notebook.
\nFinally there are three teams (see top LHS) that underwent a brutal shakedown. As @kuto0633 pointed out in the topic \"Public/Private data were sepaleted by collections\", a cursory EDA, or even simply bothering to read the collection names, one would notice that there were collections in the test data (and the Public LB evaluation) that were not in the training data. I would hazard a guess that these three teams were fitting solely to the Public LB, and didn't get the memo (I think there is a lesson to be learned in there somewhere…)

\n

Anyway, well done to everyone with what was a particularly challenging dataset.

\n

All the best,
\ncarl

", "messages": [ { "id": "1463148", "postDate": "08/10/2021 05:41:43", "content": "

Here is the shakeup for the Google Smartphone Decimeter Challenge (produced using the Shakeup interactive scatterplot maker. Right click to see an enlarged version, or see the scatterplot maker to play with an interactive version)

\n

\"\"

\n

As suggested by @t88take in the topic \"Shake up or not shake up?\" things were not too bad in the end. There were a few notable exceptions; there is a cloud of shakedowns whose public score was in the range 40-120, that dropped down to the 300-400 range. I suspect these were submissions of over-fitted ensembles. There are a good number of isoscore strings formed by people all submitting exactly the same public notebook, for example 31 teams went with the [GSDC] Smart Ensembling notebook.
\nFinally there are three teams (see top LHS) that underwent a brutal shakedown. As @kuto0633 pointed out in the topic \"Public/Private data were sepaleted by collections\", a cursory EDA, or even simply bothering to read the collection names, one would notice that there were collections in the test data (and the Public LB evaluation) that were not in the training data. I would hazard a guess that these three teams were fitting solely to the Public LB, and didn't get the memo (I think there is a lesson to be learned in there somewhere…)

\n

Anyway, well done to everyone with what was a particularly challenging dataset.

\n

All the best,
\ncarl

", "rawMarkdown": "Here is the shakeup for the Google Smartphone Decimeter Challenge (produced using the [Shakeup interactive scatterplot maker](https://www.kaggle.com/carlmcbrideellis/shakeup-interactive-scatterplot-maker?scriptVersionId=71123768). Right click to see an enlarged version, or see the scatterplot maker to play with an interactive version)\n\n![](https://raw.githubusercontent.com/Carl-McBride-Ellis/images_for_kaggle/main/Smartphone_shakeup.png)\n\nAs suggested by @t88take in the topic [\"*Shake up or not shake up?*\"](https://www.kaggle.com/c/google-smartphone-decimeter-challenge/discussion/256413) things were not too bad in the end. There were a few notable exceptions; there is a cloud of shakedowns whose public score was in the range 40-120, that dropped down to the 300-400 range. I suspect these were submissions of over-fitted ensembles. There are a good number of [isoscore strings](https://www.kaggle.com/carlmcbrideellis/shakeup-scatterplots-boxes-strings-and-things) formed by people all submitting exactly the same public notebook, for example 31 teams went with the [[GSDC] Smart Ensembling](https://www.kaggle.com/somayyehgholami/gsdc-smart-ensembling) notebook.\nFinally there are three teams (see top LHS) that underwent a brutal shakedown. As @kuto0633 pointed out in the topic [\"*Public/Private data were sepaleted by collections*\"](https://www.kaggle.com/c/google-smartphone-decimeter-challenge/discussion/250977), a cursory EDA, or even simply bothering to read the collection names, one would notice that there were collections in the test data (and the Public LB evaluation) that were not in the training data. I would hazard a guess that these three teams were fitting solely to the Public LB, and didn't get the memo (I think there is a lesson to be learned in there somewhere...)\n\nAnyway, well done to everyone with what was a particularly challenging dataset.\n\nAll the best,\ncarl", "votes": null }, { "id": "1469480", "postDate": "08/13/2021 00:52:09", "content": "

Hi Carl,

\n

I've just passed here to give you Congrats. I' m very glad with your achievement.

\n

All the best unique GM Carl McBride Ellis.

", "rawMarkdown": "Hi Carl, \n\nI've just passed here to give you Congrats. I' m very glad with your achievement.\n\nAll the best unique GM Carl McBride Ellis.", "votes": null }, { "id": "1469713", "postDate": "08/13/2021 05:08:01", "content": "

Dear @mpwolke

\n

Many thanks; your kind words now, as always, are very much appreciated!

\n

All the best,
\ncarl

", "rawMarkdown": "Dear @mpwolke \n\nMany thanks; your kind words now, as always, are very much appreciated!\n\nAll the best,\ncarl", "votes": null } ], "comments": [ { "id": 1469480, "author_name": "mpwolke", "author_url": "", "post_date": "08/13/2021 00:52:09", "content": "

Hi Carl,

\n

I've just passed here to give you Congrats. I' m very glad with your achievement.

\n

All the best unique GM Carl McBride Ellis.

", "votes": null, "replies": [ { "id": 1469713, "author_name": "carlmcbrideellis", "author_url": "", "post_date": "08/13/2021 05:08:01", "content": "

Dear @mpwolke

\n

Many thanks; your kind words now, as always, are very much appreciated!

\n

All the best,
\ncarl

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1463148": "Here is the shakeup for the Google Smartphone Decimeter Challenge (produced using the [Shakeup interactive scatterplot maker](https://www.kaggle.com/carlmcbrideellis/shakeup-interactive-scatterplot-maker?scriptVersionId=71123768). Right click to see an enlarged version, or see the scatterplot maker to play with an interactive version)\n\n![](https://raw.githubusercontent.com/Carl-McBride-Ellis/images_for_kaggle/main/Smartphone_shakeup.png)\n\nAs suggested by @t88take in the topic [\"*Shake up or not shake up?*\"](https://www.kaggle.com/c/google-smartphone-decimeter-challenge/discussion/256413) things were not too bad in the end. There were a few notable exceptions; there is a cloud of shakedowns whose public score was in the range 40-120, that dropped down to the 300-400 range. I suspect these were submissions of over-fitted ensembles. There are a good number of [isoscore strings](https://www.kaggle.com/carlmcbrideellis/shakeup-scatterplots-boxes-strings-and-things) formed by people all submitting exactly the same public notebook, for example 31 teams went with the [[GSDC] Smart Ensembling](https://www.kaggle.com/somayyehgholami/gsdc-smart-ensembling) notebook.\nFinally there are three teams (see top LHS) that underwent a brutal shakedown. As @kuto0633 pointed out in the topic [\"*Public/Private data were sepaleted by collections*\"](https://www.kaggle.com/c/google-smartphone-decimeter-challenge/discussion/250977), a cursory EDA, or even simply bothering to read the collection names, one would notice that there were collections in the test data (and the Public LB evaluation) that were not in the training data. I would hazard a guess that these three teams were fitting solely to the Public LB, and didn't get the memo (I think there is a lesson to be learned in there somewhere...)\n\nAnyway, well done to everyone with what was a particularly challenging dataset.\n\nAll the best,\ncarl", "1469480": "Hi Carl, \n\nI've just passed here to give you Congrats. I' m very glad with your achievement.\n\nAll the best unique GM Carl McBride Ellis.", "1469713": "Dear @mpwolke \n\nMany thanks; your kind words now, as always, are very much appreciated!\n\nAll the best,\ncarl" }, "source": "meta" }