many series have apparent events that are not labeled.
\nExample #1: series_id=03d92c9f6f8a. In 'train_events' there are events for nights 16-37, but they have missing step/timestamp, so they are unlabeled and are excluded from evaluation.
\nThis is easy to identify - in those nights sensor data is exactly the same as 24 hours earlier. I guess organizers filled-in bad data with previous day's data (Why? I think leaving bad data in place and letting us figure out that data is bad would be a better approach).
\nExample #2: series_id=05e1944c3818. In 'train_events' there are events for nights 2-8, but the data continues for 23 nights! All my models predict events for each of the 23 nights, and manual review of the data does not show any patterns that would explain why nights 9+ are excluded from labeled events.
\nSo far i don't have explanation for why events in nights 9+ are unlabeled. Has anyone found a way to identify situations like this? By heuristic rules, or from model predictions? There are more series like this one, so this is not an isolated situation.
", "messages": [ { "id": 2529777, "postDate": "2023-11-18T14:33:20.133Z", "content": "many series have apparent events that are not labeled.
\nExample #1: series_id=03d92c9f6f8a. In 'train_events' there are events for nights 16-37, but they have missing step/timestamp, so they are unlabeled and are excluded from evaluation.
\nThis is easy to identify - in those nights sensor data is exactly the same as 24 hours earlier. I guess organizers filled-in bad data with previous day's data (Why? I think leaving bad data in place and letting us figure out that data is bad would be a better approach).
\nExample #2: series_id=05e1944c3818. In 'train_events' there are events for nights 2-8, but the data continues for 23 nights! All my models predict events for each of the 23 nights, and manual review of the data does not show any patterns that would explain why nights 9+ are excluded from labeled events.
\nSo far i don't have explanation for why events in nights 9+ are unlabeled. Has anyone found a way to identify situations like this? By heuristic rules, or from model predictions? There are more series like this one, so this is not an isolated situation.
", "rawMarkdown": "many series have apparent events that are not labeled.\n\nExample #1: series_id=03d92c9f6f8a. In 'train_events' there are events for nights 16-37, but they have missing step/timestamp, so they are unlabeled and are excluded from evaluation.\n\n*This is easy to identify* - in those nights sensor data is exactly the same as 24 hours earlier. I guess organizers filled-in bad data with previous day's data (Why? I think leaving bad data in place and letting us figure out that data is bad would be a better approach).\n\n\nExample #2: series_id=05e1944c3818. In 'train_events' there are events for nights 2-8, but the data continues for 23 nights! All my models predict events for each of the 23 nights, and manual review of the data does not show any patterns that would explain why nights 9+ are excluded from labeled events.\n\n*So far i don't have explanation for why events in nights 9+ are unlabeled*. Has anyone found a way to identify situations like this? By heuristic rules, or from model predictions? There are more series like this one, so this is not an isolated situation.", "votes": 20 }, { "id": 2530790, "postDate": "2023-11-19T14:20:48.320Z", "content": "This does impact the results. I see 7 series like this; if i cut off predictions for them after the last true event, my CV increases by around 0.014. Maybe this is the \"magic\" in this competition?
", "rawMarkdown": "This does impact the results. I see 7 series like this; if i cut off predictions for them after the last true event, my CV increases by around 0.014. Maybe this is the \"magic\" in this competition?" }, { "id": 2529961, "postDate": "2023-11-18T17:22:39.683Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 2530114, "postDate": "2023-11-18T19:38:12.817Z", "content": "@cody11null they mentioned about eventless series not eventless days. https://www.kaggle.com/competitions/child-mind-institute-detect-sleep-states/discussion/454071
", "rawMarkdown": "@cody11null they mentioned about **eventless series** not eventless days. https://www.kaggle.com/competitions/child-mind-institute-detect-sleep-states/discussion/454071", "votes": 3 } ] } ], "comments": [ { "id": 2530790, "author_name": "Youri Matiounine", "author_url": "", "post_date": "2023-11-19T14:20:48.320000", "content": "This does impact the results. I see 7 series like this; if i cut off predictions for them after the last true event, my CV increases by around 0.014. Maybe this is the \"magic\" in this competition?
", "votes": 0, "replies": [] }, { "id": 2529961, "author_name": "", "author_url": "", "post_date": "2023-11-18T17:22:39.683000", "content": "", "votes": 0, "replies": [ { "id": 2530114, "author_name": "Anand Tirkey", "author_url": "", "post_date": "2023-11-18T19:38:12.817000", "content": "@cody11null they mentioned about eventless series not eventless days. https://www.kaggle.com/competitions/child-mind-institute-detect-sleep-states/discussion/454071
", "votes": 3, "replies": [] } ] } ], "raw_markdown_by_id": { "2529777": "many series have apparent events that are not labeled.\n\nExample #1: series_id=03d92c9f6f8a. In 'train_events' there are events for nights 16-37, but they have missing step/timestamp, so they are unlabeled and are excluded from evaluation.\n\n*This is easy to identify* - in those nights sensor data is exactly the same as 24 hours earlier. I guess organizers filled-in bad data with previous day's data (Why? I think leaving bad data in place and letting us figure out that data is bad would be a better approach).\n\n\nExample #2: series_id=05e1944c3818. In 'train_events' there are events for nights 2-8, but the data continues for 23 nights! All my models predict events for each of the 23 nights, and manual review of the data does not show any patterns that would explain why nights 9+ are excluded from labeled events.\n\n*So far i don't have explanation for why events in nights 9+ are unlabeled*. Has anyone found a way to identify situations like this? By heuristic rules, or from model predictions? There are more series like this one, so this is not an isolated situation.", "2530790": "This does impact the results. I see 7 series like this; if i cut off predictions for them after the last true event, my CV increases by around 0.014. Maybe this is the \"magic\" in this competition?", "2529961": "" } }