\nThe structure of the submission would be something like this:
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3147995%2F526bb513ab0a515ed182ad052745a138%2FCaptura%20desde%202023-10-17%2023-32-15.png?generation=1697578544042154&alt=media\")
I'm confused. Is it correct that the train data contains many days of information for each person. And the test data only contains 12 minutes of data for each test person? How can we predict onset and wakeup with only 12 minutes of data? (I assume that each person sleeps for more than 12 minutes).
I'm confused. Is it correct that the train data contains many days of information for each person. And the test data only contains 12 minutes of data for each test person? How can we predict onset and wakeup with only 12 minutes of data? (I assume that each person sleeps for more than 12 minutes).
I think that is just dummy test data, they will switch it when you submit it.
\nBase on how long my submission is taking, I think the private + public test set is around the same size as the training one.
It is a code competition so when you launch the submission your code is executed against the hidden test where there are around 200 series of a duration similar to those of the training. I can confirm that each series lasts several days.
", "rawMarkdown": "It is a code competition so when you launch the submission your code is executed against the hidden test where there are around 200 series of a duration similar to those of the training. I can confirm that each series lasts several days.", "votes": null }, { "id": "2486364", "postDate": "10/17/2023 21:03:42", "content": "Thanks @trasibulo . If each test last more than 1 day, how do we predict the onset and wakeup? I assume we should only predict 1 onset and 1 wakeup per day? Do we only use the first 24 hours of test data per person?
Because you have to predict one onset and one wakeup per day (approximately), imagine it like a segmentation problem where you have to guess the onset and the wakeup every 24 hours.
\nThat is to say, if a series has 7 days, the normal thing is that you detect 7 onsets and 7 wakeups, although there are days when the person may have taken off the watch and therefore you should not detect anything for these days (that is why I said \"approximately\").
", "rawMarkdown": "Because you have to predict one onset and one wakeup per day (approximately), imagine it like a segmentation problem where you have to guess the onset and the wakeup every 24 hours.\n\nThat is to say, if a series has 7 days, the normal thing is that you detect 7 onsets and 7 wakeups, although there are days when the person may have taken off the watch and therefore you should not detect anything for these days (that is why I said \"approximately\").", "votes": null }, { "id": "2486369", "postDate": "10/17/2023 21:14:06", "content": "Also, this is more of a change point detection kind of problem, not forecasting problem, you can use the data of that day to predict the onset and wakeup event on that day. No need to restrict yourself to predicting the future onset or wakeup events based solely on the first day's data.
", "rawMarkdown": "Also, this is more of a change point detection kind of problem, not forecasting problem, you can use the data of that day to predict the onset and wakeup event on that day. No need to restrict yourself to predicting the future onset or wakeup events based solely on the first day's data.", "votes": null }, { "id": "2486387", "postDate": "10/17/2023 21:30:13", "content": "@trasibulo Ah, so in our submission.csv file we can make multiple onset and multiple wakeup for each series_id ?
(so it seems that part of the competition is determining how many onsets each series_id has. Does it hurt our score to predict 10 onsets when a series_id only has 8 onsets?)
", "rawMarkdown": "trasibulo Ah, so in our submission.csv file we can make multiple `onset` and multiple `wakeup` for each `series_id` ?\n\n(so it seems that part of the competition is determining how many onsets each series_id has. Does it hurt our score to predict 10 onsets when a series_id only has 8 onsets?)", "votes": null }, { "id": "2486391", "postDate": "10/17/2023 21:36:21", "content": "Exactly for each night you should predict the exact time for onset and the exact time for wakeup. The more precise you are the better.
\nThe structure of the submission would be something like this:
\n
Where we see for user \"038441c925bb\" that every night we predict the onset and wakeup in a specific step
", "rawMarkdown": "Exactly for each night you should predict the exact time for onset and the exact time for wakeup. The more precise you are the better.\nThe structure of the submission would be something like this:\n\n\nWhere we see for user \"038441c925bb\" that every night we predict the onset and wakeup in a specific step", "votes": null }, { "id": "2486394", "postDate": "10/17/2023 21:43:09", "content": "Awesome. Thanks for the clarification. Now i understand.
", "rawMarkdown": "Awesome. Thanks for the clarification. Now i understand.", "votes": null }, { "id": "2486409", "postDate": "10/17/2023 22:10:54", "content": "As was discussed here, the train and test data are similar.
\n\n", "rawMarkdown": "As was discussed [here](https://www.kaggle.com/competitions/child-mind-institute-detect-sleep-states/discussion/437567#2429299), the train and test data are similar.\n\n> In general, the series in the test set will have events for more than one night, just like those in the training set.", "votes": null }, { "id": "2487471", "postDate": "10/18/2023 15:54:16", "content": "In general, the series in the test set will have events for more than one night, just like those in the training set.
\n
@trasibulo @cdeotte what if a person goes to sleep at, for example, 23:00 on a given day and wakes up at 08:00 of the next day? if we slice the test data on day intervals of 00:00 to 23:59 (of the same day) we would \"cut\" the sleeping period. Should we take sliding windows?
", "rawMarkdown": "trasibulo @cdeotte what if a person goes to sleep at, for example, 23:00 on a given day and wakes up at 08:00 of the next day? if we slice the test data on day intervals of 00:00 to 23:59 (of the same day) we would \"cut\" the sleeping period. Should we take sliding windows?", "votes": null }, { "id": "2487514", "postDate": "10/18/2023 16:10:44", "content": "The data page says:
\n\n\nSleep events do not need to straddle the day-line, and therefore there is no hard rule defining how many may occur within a given period. However, no more than one window should be assigned per night. For example, it is valid for an individual to have a sleep window from 01h00–06h00 and 19h00–23h30 in the same calendar day, though assigned to consecutive nights
\n
So i assume that yes we must use a sliding technique to find up to 1 window per night.
", "rawMarkdown": "The data page says:\n>Sleep events do not need to straddle the day-line, and therefore there is no hard rule defining how many may occur within a given period. However, no more than one window should be assigned per night. For example, it is valid for an individual to have a sleep window from 01h00–06h00 and 19h00–23h30 in the same calendar day, though assigned to consecutive nights\n\nSo i assume that yes we must use a sliding technique to find up to 1 window per night.", "votes": null }, { "id": "2493098", "postDate": "10/23/2023 06:47:53", "content": "@cdeotte The statement that the LB test data only contains 12 minutes of data per person is incorrect. The LB (Leaderboard) test data refers to the test dataset used in machine learning competitions, where participants submit their predictions for evaluation. The specific details of the test data for a particular competition can vary, but it typically consists of a set of features or inputs for each person, and participants are required to predict certain outcomes or targets based on these inputs.
\nThe duration of the data provided in the test set can vary depending on the competition and the specific problem being addressed. It is not necessarily limited to 12 minutes of data per person. The duration of the data could be longer or shorter, depending on the problem and the available data.
\nRegarding the prediction of onset and wakeup with only 12 minutes of data, it is important to note that machine learning models can learn patterns and make predictions based on the available data. While it is true that a person's sleep duration is typically longer than 12 minutes, it is possible for machine learning models to extract relevant features from the given data and make predictions based on those features. The models can learn patterns and relationships between the provided data and the target variables, allowing them to make predictions even with limited data.
", "rawMarkdown": "cdeotte The statement that the LB test data only contains 12 minutes of data per person is incorrect. The LB (Leaderboard) test data refers to the test dataset used in machine learning competitions, where participants submit their predictions for evaluation. The specific details of the test data for a particular competition can vary, but it typically consists of a set of features or inputs for each person, and participants are required to predict certain outcomes or targets based on these inputs.\n\nThe duration of the data provided in the test set can vary depending on the competition and the specific problem being addressed. It is not necessarily limited to 12 minutes of data per person. The duration of the data could be longer or shorter, depending on the problem and the available data.\n\nRegarding the prediction of onset and wakeup with only 12 minutes of data, it is important to note that machine learning models can learn patterns and make predictions based on the available data. While it is true that a person's sleep duration is typically longer than 12 minutes, it is possible for machine learning models to extract relevant features from the given data and make predictions based on those features. The models can learn patterns and relationships between the provided data and the target variables, allowing them to make predictions even with limited data.", "votes": null }, { "id": "2497228", "postDate": "10/24/2023 14:27:23", "content": "Sorry, let me clarify. I still don't get it.
\nassigned to consecutive nights means \"the sleep from 10/22 22:00 to 10/23 06:00 is divided into 22:00-24:00 and 0:00-06:00\", doesn't it?
No, my understanding of \"night\" is between Monday noon 12p to Tuesday noon 12p. (So a single night is half in one day and half in the next day, but it is one night)
", "rawMarkdown": "No, my understanding of \"night\" is between Monday **noon 12p** to Tuesday **noon 12p**. (So a single night is half in one day and half in the next day, but it is **one night**)", "votes": null }, { "id": "2498613", "postDate": "10/25/2023 12:17:22", "content": "I got it. Thanks Chris!
", "rawMarkdown": "I got it. Thanks Chris!", "votes": null }, { "id": "2522741", "postDate": "11/13/2023 02:42:20", "content": "If I assume \"night\" has one day interval starting from the first date of the series in local time zone and starts from noon 12 to the next 12 noon, all events happen after the start time. But, there are a few events that don't happen until the end time.
\n| series_id | \nnight | \nevent | \nlocal_time | \nnight_start_time | \nnight_end_time | \n
|---|---|---|---|---|---|
| 08db4255286f | \n23 | \nwakeup | \n2018-11-28 12:30:00 | \n2018-11-27 12:00:00 | \n2018-11-28 12:00:00 | \n
| 51fdcc8d9fe7 | \n15 | \nonset | \n2018-11-14 21:03:00 | \n2018-11-13 12:00:00 | \n2018-11-14 12:00:00 | \n
So, I think \"night\" starts from noon 12 to the time \"wakeup\" happens.
", "rawMarkdown": "If I assume \"night\" has one day interval starting from the first date of the series in local time zone and starts from noon 12 to the next 12 noon, all events happen after the start time. But, there are a few events that don't happen until the end time.\n\n| series_id | night | event | local_time | night_start_time | night_end_time |\n| --- | --- | --- | --- | --- | --- |\n| 08db4255286f | 23 | wakeup | 2018-11-28 12:30:00 | 2018-11-27 12:00:00 | 2018-11-28 12:00:00 |\n| 51fdcc8d9fe7 | 15 | onset | 2018-11-14 21:03:00 | 2018-11-13 12:00:00 | 2018-11-14 12:00:00 |\n\nSo, I think \"night\" starts from noon 12 to the time \"wakeup\" happens.", "votes": null } ], "comments": [ { "id": 2486352, "author_name": "yuchen2066", "author_url": "", "post_date": "10/17/2023 20:52:35", "content": "I think that is just dummy test data, they will switch it when you submit it.
\nBase on how long my submission is taking, I think the private + public test set is around the same size as the training one.
It is a code competition so when you launch the submission your code is executed against the hidden test where there are around 200 series of a duration similar to those of the training. I can confirm that each series lasts several days.
", "votes": null, "replies": [ { "id": 2486364, "author_name": "cdeotte", "author_url": "", "post_date": "10/17/2023 21:03:42", "content": "Thanks @trasibulo . If each test last more than 1 day, how do we predict the onset and wakeup? I assume we should only predict 1 onset and 1 wakeup per day? Do we only use the first 24 hours of test data per person?
Because you have to predict one onset and one wakeup per day (approximately), imagine it like a segmentation problem where you have to guess the onset and the wakeup every 24 hours.
\nThat is to say, if a series has 7 days, the normal thing is that you detect 7 onsets and 7 wakeups, although there are days when the person may have taken off the watch and therefore you should not detect anything for these days (that is why I said \"approximately\").
", "votes": null, "replies": [ { "id": 2486387, "author_name": "cdeotte", "author_url": "", "post_date": "10/17/2023 21:30:13", "content": "@trasibulo Ah, so in our submission.csv file we can make multiple onset and multiple wakeup for each series_id ?
(so it seems that part of the competition is determining how many onsets each series_id has. Does it hurt our score to predict 10 onsets when a series_id only has 8 onsets?)
", "votes": null, "replies": [ { "id": 2486391, "author_name": "trasibulo", "author_url": "", "post_date": "10/17/2023 21:36:21", "content": "Exactly for each night you should predict the exact time for onset and the exact time for wakeup. The more precise you are the better.
\nThe structure of the submission would be something like this:
\n
Where we see for user \"038441c925bb\" that every night we predict the onset and wakeup in a specific step
", "votes": null, "replies": [ { "id": 2486394, "author_name": "cdeotte", "author_url": "", "post_date": "10/17/2023 21:43:09", "content": "Awesome. Thanks for the clarification. Now i understand.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2486369, "author_name": "yuchen2066", "author_url": "", "post_date": "10/17/2023 21:14:06", "content": "Also, this is more of a change point detection kind of problem, not forecasting problem, you can use the data of that day to predict the onset and wakeup event on that day. No need to restrict yourself to predicting the future onset or wakeup events based solely on the first day's data.
", "votes": null, "replies": [] } ] }, { "id": 2487471, "author_name": "alejopaullier", "author_url": "", "post_date": "10/18/2023 15:54:16", "content": "@trasibulo @cdeotte what if a person goes to sleep at, for example, 23:00 on a given day and wakes up at 08:00 of the next day? if we slice the test data on day intervals of 00:00 to 23:59 (of the same day) we would \"cut\" the sleeping period. Should we take sliding windows?
", "votes": null, "replies": [ { "id": 2487514, "author_name": "cdeotte", "author_url": "", "post_date": "10/18/2023 16:10:44", "content": "The data page says:
\n\n\nSleep events do not need to straddle the day-line, and therefore there is no hard rule defining how many may occur within a given period. However, no more than one window should be assigned per night. For example, it is valid for an individual to have a sleep window from 01h00–06h00 and 19h00–23h30 in the same calendar day, though assigned to consecutive nights
\n
So i assume that yes we must use a sliding technique to find up to 1 window per night.
", "votes": null, "replies": [ { "id": 2497228, "author_name": "econjt", "author_url": "", "post_date": "10/24/2023 14:27:23", "content": "Sorry, let me clarify. I still don't get it.
\nassigned to consecutive nights means \"the sleep from 10/22 22:00 to 10/23 06:00 is divided into 22:00-24:00 and 0:00-06:00\", doesn't it?
No, my understanding of \"night\" is between Monday noon 12p to Tuesday noon 12p. (So a single night is half in one day and half in the next day, but it is one night)
", "votes": null, "replies": [ { "id": 2498613, "author_name": "econjt", "author_url": "", "post_date": "10/25/2023 12:17:22", "content": "I got it. Thanks Chris!
", "votes": null, "replies": [] }, { "id": 2522741, "author_name": "dongyk", "author_url": "", "post_date": "11/13/2023 02:42:20", "content": "If I assume \"night\" has one day interval starting from the first date of the series in local time zone and starts from noon 12 to the next 12 noon, all events happen after the start time. But, there are a few events that don't happen until the end time.
\n| series_id | \nnight | \nevent | \nlocal_time | \nnight_start_time | \nnight_end_time | \n
|---|---|---|---|---|---|
| 08db4255286f | \n23 | \nwakeup | \n2018-11-28 12:30:00 | \n2018-11-27 12:00:00 | \n2018-11-28 12:00:00 | \n
| 51fdcc8d9fe7 | \n15 | \nonset | \n2018-11-14 21:03:00 | \n2018-11-13 12:00:00 | \n2018-11-14 12:00:00 | \n
So, I think \"night\" starts from noon 12 to the time \"wakeup\" happens.
", "votes": null, "replies": [] } ] } ] } ] } ] } ] }, { "id": 2486409, "author_name": "mitsuyasuhoshino", "author_url": "", "post_date": "10/17/2023 22:10:54", "content": "As was discussed here, the train and test data are similar.
\n\n", "votes": null, "replies": [] }, { "id": 2493098, "author_name": "muhammadqasimiqbal", "author_url": "", "post_date": "10/23/2023 06:47:53", "content": "In general, the series in the test set will have events for more than one night, just like those in the training set.
\n
@cdeotte The statement that the LB test data only contains 12 minutes of data per person is incorrect. The LB (Leaderboard) test data refers to the test dataset used in machine learning competitions, where participants submit their predictions for evaluation. The specific details of the test data for a particular competition can vary, but it typically consists of a set of features or inputs for each person, and participants are required to predict certain outcomes or targets based on these inputs.
\nThe duration of the data provided in the test set can vary depending on the competition and the specific problem being addressed. It is not necessarily limited to 12 minutes of data per person. The duration of the data could be longer or shorter, depending on the problem and the available data.
\nRegarding the prediction of onset and wakeup with only 12 minutes of data, it is important to note that machine learning models can learn patterns and make predictions based on the available data. While it is true that a person's sleep duration is typically longer than 12 minutes, it is possible for machine learning models to extract relevant features from the given data and make predictions based on those features. The models can learn patterns and relationships between the provided data and the target variables, allowing them to make predictions even with limited data.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2486346": "I'm confused. Is it correct that the train data contains many days of information for each person. And the test data only contains 12 minutes of data for each test person? How can we predict `onset` and `wakeup` with only 12 minutes of data? (I assume that each person sleeps for more than 12 minutes).", "2486352": "I think that is just dummy test data, they will switch it when you submit it. \nBase on how long my submission is taking, I think the private + public test set is around the same size as the training one.", "2486359": "It is a code competition so when you launch the submission your code is executed against the hidden test where there are around 200 series of a duration similar to those of the training. I can confirm that each series lasts several days.", "2486364": "Thanks @trasibulo . If each test last more than 1 day, how do we predict the `onset` and `wakeup`? I assume we should only predict 1 onset and 1 wakeup per day? Do we only use the first 24 hours of test data per person?", "2486366": "Because you have to predict one onset and one wakeup per day (approximately), imagine it like a segmentation problem where you have to guess the onset and the wakeup every 24 hours.\n\nThat is to say, if a series has 7 days, the normal thing is that you detect 7 onsets and 7 wakeups, although there are days when the person may have taken off the watch and therefore you should not detect anything for these days (that is why I said \"approximately\").", "2486369": "Also, this is more of a change point detection kind of problem, not forecasting problem, you can use the data of that day to predict the onset and wakeup event on that day. No need to restrict yourself to predicting the future onset or wakeup events based solely on the first day's data.", "2486387": "trasibulo Ah, so in our submission.csv file we can make multiple `onset` and multiple `wakeup` for each `series_id` ?\n\n(so it seems that part of the competition is determining how many onsets each series_id has. Does it hurt our score to predict 10 onsets when a series_id only has 8 onsets?)", "2486391": "Exactly for each night you should predict the exact time for onset and the exact time for wakeup. The more precise you are the better.\nThe structure of the submission would be something like this:\n\n\nWhere we see for user \"038441c925bb\" that every night we predict the onset and wakeup in a specific step", "2486394": "Awesome. Thanks for the clarification. Now i understand.", "2486409": "As was discussed [here](https://www.kaggle.com/competitions/child-mind-institute-detect-sleep-states/discussion/437567#2429299), the train and test data are similar.\n\n> In general, the series in the test set will have events for more than one night, just like those in the training set.", "2487471": "trasibulo @cdeotte what if a person goes to sleep at, for example, 23:00 on a given day and wakes up at 08:00 of the next day? if we slice the test data on day intervals of 00:00 to 23:59 (of the same day) we would \"cut\" the sleeping period. Should we take sliding windows?", "2487514": "The data page says:\n>Sleep events do not need to straddle the day-line, and therefore there is no hard rule defining how many may occur within a given period. However, no more than one window should be assigned per night. For example, it is valid for an individual to have a sleep window from 01h00–06h00 and 19h00–23h30 in the same calendar day, though assigned to consecutive nights\n\nSo i assume that yes we must use a sliding technique to find up to 1 window per night.", "2493098": "cdeotte The statement that the LB test data only contains 12 minutes of data per person is incorrect. The LB (Leaderboard) test data refers to the test dataset used in machine learning competitions, where participants submit their predictions for evaluation. The specific details of the test data for a particular competition can vary, but it typically consists of a set of features or inputs for each person, and participants are required to predict certain outcomes or targets based on these inputs.\n\nThe duration of the data provided in the test set can vary depending on the competition and the specific problem being addressed. It is not necessarily limited to 12 minutes of data per person. The duration of the data could be longer or shorter, depending on the problem and the available data.\n\nRegarding the prediction of onset and wakeup with only 12 minutes of data, it is important to note that machine learning models can learn patterns and make predictions based on the available data. While it is true that a person's sleep duration is typically longer than 12 minutes, it is possible for machine learning models to extract relevant features from the given data and make predictions based on those features. The models can learn patterns and relationships between the provided data and the target variables, allowing them to make predictions even with limited data.", "2497228": "Sorry, let me clarify. I still don't get it.\n`assigned to consecutive nights` means \"the sleep from 10/22 22:00 to 10/23 06:00 is divided into 22:00-24:00 and 0:00-06:00\", doesn't it?", "2497287": "No, my understanding of \"night\" is between Monday **noon 12p** to Tuesday **noon 12p**. (So a single night is half in one day and half in the next day, but it is **one night**)", "2498613": "I got it. Thanks Chris!", "2522741": "If I assume \"night\" has one day interval starting from the first date of the series in local time zone and starts from noon 12 to the next 12 noon, all events happen after the start time. But, there are a few events that don't happen until the end time.\n\n| series_id | night | event | local_time | night_start_time | night_end_time |\n| --- | --- | --- | --- | --- | --- |\n| 08db4255286f | 23 | wakeup | 2018-11-28 12:30:00 | 2018-11-27 12:00:00 | 2018-11-28 12:00:00 |\n| 51fdcc8d9fe7 | 15 | onset | 2018-11-14 21:03:00 | 2018-11-13 12:00:00 | 2018-11-14 12:00:00 |\n\nSo, I think \"night\" starts from noon 12 to the time \"wakeup\" happens." }, "source": "meta" }