{ "id": 580988, "title": "There is a leak in train files", "url": "/competitions/waveform-inversion/discussion/580988", "author_name": "", "post_date": "2025-05-27T16:47:01.182060400Z", "votes": 30, "comment_count": 6, "views": 0, "content": "

The leak: for CurveFault_B and FlatFault_B, and only for them, the last row in seis equals zero.
\nI did not check for all the files. But it was true for all the files I checked.
\nI did not see this leak in the test files, so it's probably only in train.
\nConsequence: Since the leak is only in train files, train and infer the models without the last row. You can also ignore it; it's a minor one.
\nLet's hope this is the only leak…

\n

Demonstration of the leak here.

", "messages": [ { "id": "3210759", "postDate": "05/27/2025 16:47:01", "content": "

The leak: for CurveFault_B and FlatFault_B, and only for them, the last row in seis equals zero.
\nI did not check for all the files. But it was true for all the files I checked.
\nI did not see this leak in the test files, so it's probably only in train.
\nConsequence: Since the leak is only in train files, train and infer the models without the last row. You can also ignore it; it's a minor one.
\nLet's hope this is the only leak…

\n

Demonstration of the leak here.

", "rawMarkdown": "The leak: for CurveFault_B and FlatFault_B, and only for them, the last row in seis equals zero.\nI did not check for all the files. But it was true for all the files I checked. \nI did not see this leak in the test files, so it's probably only in train.\nConsequence: Since the leak is only in train files, train and infer the models without the last row. You can also ignore it; it's a minor one.\nLet's hope this is the only leak...\n\nDemonstration of the leak [here](https://www.kaggle.com/code/shlomoron/gwi-leak).", "votes": null }, { "id": "3210825", "postDate": "05/27/2025 17:52:28", "content": "

Nice find! I believe the responses at these later times are mostly affected by the boundary condition and hence probably aren't really informative anyways.

", "rawMarkdown": "Nice find! I believe the responses at these later times are mostly affected by the boundary condition and hence probably aren't really informative anyways.", "votes": null }, { "id": "3211106", "postDate": "05/28/2025 04:59:27", "content": "

Did you do something analysis on how much the score changes by training with and without that last row?

", "rawMarkdown": "Did you do something analysis on how much the score changes by training with and without that last row?", "votes": null }, { "id": "3211219", "postDate": "05/28/2025 07:51:27", "content": "

I did not.

", "rawMarkdown": "I did not.", "votes": null }, { "id": "3211378", "postDate": "05/28/2025 11:02:58", "content": "

Thanks for letting us know but I don't understand the terminology here.

\n

Why is that called a leak? Is that a different concept than this?

", "rawMarkdown": "Thanks for letting us know but I don't understand the terminology here. \n\nWhy is that called a leak? Is that a different concept than [this](https://en.wikipedia.org/wiki/Leakage_(machine_learning))?", "votes": null }, { "id": "3211421", "postDate": "05/28/2025 12:09:13", "content": "

Here you go:
\nhttps://youtu.be/sEJHyuWKd-s?t=412

\n

Here greySnow is mentioning that if the last row of a given seismic sample is zero, then the small leak is that the sample must be either CurveFault_B or FlatFault_B

", "rawMarkdown": "Here you go:\nhttps://youtu.be/sEJHyuWKd-s?t=412\n\nHere greySnow is mentioning that if the last row of a given seismic sample is zero, then the small leak is that the sample must be either CurveFault_B or FlatFault_B", "votes": null }, { "id": "3211456", "postDate": "05/28/2025 13:06:01", "content": "

Oh, I see. Thanks!

", "rawMarkdown": "Oh, I see. Thanks!", "votes": null } ], "comments": [ { "id": 3210825, "author_name": "hoffmanns", "author_url": "", "post_date": "05/27/2025 17:52:28", "content": "

Nice find! I believe the responses at these later times are mostly affected by the boundary condition and hence probably aren't really informative anyways.

", "votes": null, "replies": [] }, { "id": 3211106, "author_name": "fpeccia", "author_url": "", "post_date": "05/28/2025 04:59:27", "content": "

Did you do something analysis on how much the score changes by training with and without that last row?

", "votes": null, "replies": [ { "id": 3211219, "author_name": "shlomoron", "author_url": "", "post_date": "05/28/2025 07:51:27", "content": "

I did not.

", "votes": null, "replies": [] } ] }, { "id": 3211378, "author_name": "mgoksu", "author_url": "", "post_date": "05/28/2025 11:02:58", "content": "

Thanks for letting us know but I don't understand the terminology here.

\n

Why is that called a leak? Is that a different concept than this?

", "votes": null, "replies": [ { "id": 3211421, "author_name": "harshitsheoran", "author_url": "", "post_date": "05/28/2025 12:09:13", "content": "

Here you go:
\nhttps://youtu.be/sEJHyuWKd-s?t=412

\n

Here greySnow is mentioning that if the last row of a given seismic sample is zero, then the small leak is that the sample must be either CurveFault_B or FlatFault_B

", "votes": null, "replies": [ { "id": 3211456, "author_name": "mgoksu", "author_url": "", "post_date": "05/28/2025 13:06:01", "content": "

Oh, I see. Thanks!

", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3210759": "The leak: for CurveFault_B and FlatFault_B, and only for them, the last row in seis equals zero.\nI did not check for all the files. But it was true for all the files I checked. \nI did not see this leak in the test files, so it's probably only in train.\nConsequence: Since the leak is only in train files, train and infer the models without the last row. You can also ignore it; it's a minor one.\nLet's hope this is the only leak...\n\nDemonstration of the leak [here](https://www.kaggle.com/code/shlomoron/gwi-leak).", "3210825": "Nice find! I believe the responses at these later times are mostly affected by the boundary condition and hence probably aren't really informative anyways.", "3211106": "Did you do something analysis on how much the score changes by training with and without that last row?", "3211219": "I did not.", "3211378": "Thanks for letting us know but I don't understand the terminology here. \n\nWhy is that called a leak? Is that a different concept than [this](https://en.wikipedia.org/wiki/Leakage_(machine_learning))?", "3211421": "Here you go:\nhttps://youtu.be/sEJHyuWKd-s?t=412\n\nHere greySnow is mentioning that if the last row of a given seismic sample is zero, then the small leak is that the sample must be either CurveFault_B or FlatFault_B", "3211456": "Oh, I see. Thanks!" }, "source": "meta" }