{ "id": 228898, "title": "Raw timestamps cause leakage?", "url": "/competitions/indoor-location-navigation/discussion/228898", "author_name": "", "post_date": "2021-03-27T03:52:10.066680800Z", "votes": 47, "comment_count": 2, "views": 0, "content": "

I am sorry if I bother you so much.
\nIn training dataset, I can find that the last waypoint of a path is exactly the same as the first waypoint of another path.
\nI think this is because these data are divided from a single measurement.
\nI wonder whether the host considered this point when creating test dataset, which could cause a leakage.
\nAlthough start time of each test data is set to 0, we can retrieve raw timestamp with some tricks.

\n

The picture is created with getting the first and last waypoints and just sorting the data with their timestamps.
\n\"https://www.kaggle.com/tomooinubushi/pictures?select=Screenshot_2021-03-27+EDA+for+indoor+Kaggle.png\"

\n

I also noticed that a person reported that retrieving raw timestamp in test set improved LB score 0.4. I suspect that if the host measured the data floor by floor, the raw timestamp has some information about the floor and cause another leakage. Do you think it is true? The below figure is a boxplot of start time in one site from training data.
\n\"https://www.kaggle.com/tomooinubushi/pictures?select=__results___14_10.png\"

\n

I wonder whether the retrieved timestamps in test set cause any leakages.
\nMaybe, I am missing something.
\nAny comments are welcome!

\n

[update]I created public notebook about this leakage. I am now confident about the existence of it.

", "messages": [ { "id": "1253805", "postDate": "03/27/2021 03:52:10", "content": "

I am sorry if I bother you so much.
\nIn training dataset, I can find that the last waypoint of a path is exactly the same as the first waypoint of another path.
\nI think this is because these data are divided from a single measurement.
\nI wonder whether the host considered this point when creating test dataset, which could cause a leakage.
\nAlthough start time of each test data is set to 0, we can retrieve raw timestamp with some tricks.

\n

The picture is created with getting the first and last waypoints and just sorting the data with their timestamps.
\n\"https://www.kaggle.com/tomooinubushi/pictures?select=Screenshot_2021-03-27+EDA+for+indoor+Kaggle.png\"

\n

I also noticed that a person reported that retrieving raw timestamp in test set improved LB score 0.4. I suspect that if the host measured the data floor by floor, the raw timestamp has some information about the floor and cause another leakage. Do you think it is true? The below figure is a boxplot of start time in one site from training data.
\n\"https://www.kaggle.com/tomooinubushi/pictures?select=__results___14_10.png\"

\n

I wonder whether the retrieved timestamps in test set cause any leakages.
\nMaybe, I am missing something.
\nAny comments are welcome!

\n

[update]I created public notebook about this leakage. I am now confident about the existence of it.

", "rawMarkdown": "I am sorry if I bother you so much.\nIn training dataset, I can find that the last waypoint of a path is exactly the same as the first waypoint of another path.\nI think this is because these data are divided from a single measurement.\nI wonder whether the host considered this point when creating test dataset, which could cause a leakage.\nAlthough start time of each test data is set to 0, we can retrieve raw timestamp with [some tricks](https://www.kaggle.com/jiweiliu/fix-the-timestamps-of-test-data-using-dask).\n\nThe picture is created with getting the first and last waypoints and just sorting the data with their timestamps.\n![https://www.kaggle.com/tomooinubushi/pictures?select=Screenshot_2021-03-27+EDA+for+indoor+Kaggle.png](https://storage.googleapis.com/kagglesdsdata/datasets/1235322/2062249/Screenshot_2021-03-27%20EDA%20for%20indoor%20Kaggle.png?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=databundle-worker-v2%40kaggle-161607.iam.gserviceaccount.com%2F20210501%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20210501T101444Z&X-Goog-Expires=172799&X-Goog-SignedHeaders=host&X-Goog-Signature=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)\n\nI also noticed that [a person reported](https://www.kaggle.com/jiweiliu/fix-the-timestamps-of-test-data-using-dask/comments) that retrieving raw timestamp in test set improved LB score 0.4. I suspect that if the host measured the data floor by floor, the raw timestamp has some information about the floor and cause another leakage. Do you think it is true? The below figure is a boxplot of start time in one site from training data.\n![https://www.kaggle.com/tomooinubushi/pictures?select=__results___14_10.png](https://storage.googleapis.com/kagglesdsdata/datasets/1235322/2062249/__results___14_10.png?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=databundle-worker-v2%40kaggle-161607.iam.gserviceaccount.com%2F20210501%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20210501T101519Z&X-Goog-Expires=172799&X-Goog-SignedHeaders=host&X-Goog-Signature=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)\n\n\nI wonder whether the retrieved timestamps in test set cause any leakages.\nMaybe, I am missing something.\nAny comments are welcome!\n\n[update]I created [public notebook](https://www.kaggle.com/tomooinubushi/postprocessing-based-on-leakage) about this leakage. I am now confident about the existence of it.", "votes": null }, { "id": "1259745", "postDate": "04/01/2021 16:22:13", "content": "

yes leak.
\nI noticed this but didn't tell…
\n\"\"

", "rawMarkdown": "yes leak.\nI noticed this but didn't tell...\n![](https://pbs.twimg.com/media/Ex5sKCOVIAAYaxS?format=png&name=900x900)", "votes": null }, { "id": "1260262", "postDate": "04/02/2021 01:24:53", "content": "

Thank you for your comment.
\nX and Y axes represent raw timestamps of train and test, and dots represent site-floor-wise mean?
\nIf so, the figure is so beautiful and the leak is so strong.
\nAs the method of retrieving raw timestamps have been already opened, I thought considerable number of people had already noticed this point.

", "rawMarkdown": "Thank you for your comment.\nX and Y axes represent raw timestamps of train and test, and dots represent site-floor-wise mean?\nIf so, the figure is so beautiful and the leak is so strong.\nAs [the method of retrieving raw timestamps have been already opened](https://www.kaggle.com/jiweiliu/fix-the-timestamps-of-test-data-using-dask), I thought considerable number of people had already noticed this point.", "votes": null } ], "comments": [ { "id": 1259745, "author_name": "ryunosukeishizaki", "author_url": "", "post_date": "04/01/2021 16:22:13", "content": "

yes leak.
\nI noticed this but didn't tell…
\n\"\"

", "votes": null, "replies": [ { "id": 1260262, "author_name": "tomooinubushi", "author_url": "", "post_date": "04/02/2021 01:24:53", "content": "

Thank you for your comment.
\nX and Y axes represent raw timestamps of train and test, and dots represent site-floor-wise mean?
\nIf so, the figure is so beautiful and the leak is so strong.
\nAs the method of retrieving raw timestamps have been already opened, I thought considerable number of people had already noticed this point.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1253805": "I am sorry if I bother you so much.\nIn training dataset, I can find that the last waypoint of a path is exactly the same as the first waypoint of another path.\nI think this is because these data are divided from a single measurement.\nI wonder whether the host considered this point when creating test dataset, which could cause a leakage.\nAlthough start time of each test data is set to 0, we can retrieve raw timestamp with [some tricks](https://www.kaggle.com/jiweiliu/fix-the-timestamps-of-test-data-using-dask).\n\nThe picture is created with getting the first and last waypoints and just sorting the data with their timestamps.\n![https://www.kaggle.com/tomooinubushi/pictures?select=Screenshot_2021-03-27+EDA+for+indoor+Kaggle.png](https://storage.googleapis.com/kagglesdsdata/datasets/1235322/2062249/Screenshot_2021-03-27%20EDA%20for%20indoor%20Kaggle.png?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=databundle-worker-v2%40kaggle-161607.iam.gserviceaccount.com%2F20210501%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20210501T101444Z&X-Goog-Expires=172799&X-Goog-SignedHeaders=host&X-Goog-Signature=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)\n\nI also noticed that [a person reported](https://www.kaggle.com/jiweiliu/fix-the-timestamps-of-test-data-using-dask/comments) that retrieving raw timestamp in test set improved LB score 0.4. I suspect that if the host measured the data floor by floor, the raw timestamp has some information about the floor and cause another leakage. Do you think it is true? The below figure is a boxplot of start time in one site from training data.\n![https://www.kaggle.com/tomooinubushi/pictures?select=__results___14_10.png](https://storage.googleapis.com/kagglesdsdata/datasets/1235322/2062249/__results___14_10.png?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=databundle-worker-v2%40kaggle-161607.iam.gserviceaccount.com%2F20210501%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20210501T101519Z&X-Goog-Expires=172799&X-Goog-SignedHeaders=host&X-Goog-Signature=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)\n\n\nI wonder whether the retrieved timestamps in test set cause any leakages.\nMaybe, I am missing something.\nAny comments are welcome!\n\n[update]I created [public notebook](https://www.kaggle.com/tomooinubushi/postprocessing-based-on-leakage) about this leakage. I am now confident about the existence of it.", "1259745": "yes leak.\nI noticed this but didn't tell...\n![](https://pbs.twimg.com/media/Ex5sKCOVIAAYaxS?format=png&name=900x900)", "1260262": "Thank you for your comment.\nX and Y axes represent raw timestamps of train and test, and dots represent site-floor-wise mean?\nIf so, the figure is so beautiful and the leak is so strong.\nAs [the method of retrieving raw timestamps have been already opened](https://www.kaggle.com/jiweiliu/fix-the-timestamps-of-test-data-using-dask), I thought considerable number of people had already noticed this point." }, "source": "meta" }