![\"\"](\"https://i.ibb.co/jVTrN2M/building77.jpg\")
I noticed there are some negative readings in site 0 test set. They always occur right before a positive outlier. They don't exist in training set. For example; this is from building_id 77. (Those are not my predictions, they are from the ucf website)
\n\n
![\"\"](\"https://i.ibb.co/vjfDW9D/neg.jpg\")
What are they? Are we supposed to predict them or are they going to be excluded from rmsle calculation?
", "messages": [ { "id": 669524, "postDate": "2019-11-10T06:09:07.037Z", "content": "I noticed there are some negative readings in site 0 test set. They always occur right before a positive outlier. They don't exist in training set. For example; this is from building_id 77. (Those are not my predictions, they are from the ucf website)
\n\n
What are they? Are we supposed to predict them or are they going to be excluded from rmsle calculation?
", "rawMarkdown": "I noticed there are some negative readings in site 0 test set. They always occur right before a positive outlier. They don't exist in training set. For example; this is from building_id 77. **(Those are not my predictions, they are from the ucf website)**\n\n\n\n\n\nWhat are they? Are we supposed to predict them or are they going to be excluded from rmsle calculation? ", "votes": 30 }, { "id": 669894, "postDate": "2019-11-10T16:18:31.107Z", "content": "They seem to be very rare so they will be washed out by rmsle.
", "rawMarkdown": "They seem to be very rare so they will be washed out by rmsle.", "votes": 1 }, { "id": 669851, "postDate": "2019-11-10T15:09:27.950Z", "content": "Seriously? When are we going to get a clean competition again???
\n\nWho is putting a kernel with all the override that I will upvote for sure?
", "rawMarkdown": "Seriously? When are we going to get a clean competition again???\n\nWho is putting a kernel with all the override that I will upvote for sure?", "votes": 2 }, { "id": 669766, "postDate": "2019-11-10T12:47:16.050Z", "content": "2 submissions ;-)
", "rawMarkdown": "2 submissions ;-)", "votes": 2 }, { "id": 669688, "postDate": "2019-11-10T10:26:26.357Z", "content": "I notice the same reading also but it doesn't make sense how can meter reading be in negative value.
", "rawMarkdown": "I notice the same reading also but it doesn't make sense how can meter reading be in negative value.", "votes": 2, "replies": [ { "id": 669694, "postDate": "2019-11-10T10:34:07.900Z", "content": "My theory is they are random power outages. That's why meter suddenly outputs a huge negative number (when the electricity goes off). When it comes back, there is a huge positive number.
", "rawMarkdown": "My theory is they are random power outages. That's why meter suddenly outputs a huge negative number (when the electricity goes off). When it comes back, there is a huge positive number." }, { "id": 669737, "postDate": "2019-11-10T12:13:10.187Z", "content": "Or perhaps you just have to change sign ?
\n\nI really don’t know, but this could be an option..
\n\nBut the magnitude is strange anyway. Unless you divide it by 10 ?
", "rawMarkdown": "Or perhaps you just have to change sign ? \n\nI really don’t know, but this could be an option..\n\nBut the magnitude is strange anyway. Unless you divide it by 10 ?\n" } ] }, { "id": 681948, "postDate": "2019-11-26T18:21:07.517Z", "content": "What happens to the test performance when you remove the positive outlier and the negative row?
", "rawMarkdown": "What happens to the test performance when you remove the positive outlier and the negative row?" }, { "id": 681851, "postDate": "2019-11-26T15:59:05.313Z", "content": "The difference between them seems to be the true reading for both readings. For example (1479-1209)/2 = 135 kWh, which is the value of their closest neighbors.
", "rawMarkdown": "The difference between them seems to be the true reading for both readings. For example (1479-1209)/2 = 135 kWh, which is the value of their closest neighbors. " }, { "id": 671418, "postDate": "2019-11-12T17:09:37.950Z", "content": "You might have negative readings if power were being produced on site, but that seems unlikely in this case.
", "rawMarkdown": "You might have negative readings if power were being produced on site, but that seems unlikely in this case." }, { "id": 669789, "postDate": "2019-11-10T13:40:20.533Z", "content": "Are you talking about this data for building 77? \nhttps://www.oeis.ucf.edu/buildings/68
", "rawMarkdown": "Are you talking about this data for building 77? \nhttps://www.oeis.ucf.edu/buildings/68\n", "replies": [ { "id": 669792, "postDate": "2019-11-10T13:46:12.190Z", "content": "Nope, it is from this building.\nhttps://www.oeis.ucf.edu/buildings/131
", "rawMarkdown": "Nope, it is from this building.\nhttps://www.oeis.ucf.edu/buildings/131", "votes": 3 }, { "id": 669895, "postDate": "2019-11-10T16:28:43.073Z", "content": "Thanks, in fact I had 4 candidates for this building:\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F698363%2F452373039bf6b3659f44b0da850f5bab%2Fbuilding_77_candidates.png?generation=1573403081891215&alt=media\")
I was able to retrieve and join on [\"square_feet\", \"primary_use\", \"year_built\"] but I have a lot of buildings with two or more candidates. Did you figure out how to select the correct one? Download and compare data?
", "rawMarkdown": "Thanks, in fact I had 4 candidates for this building:\n\n\nI was able to retrieve and join on [\"square_feet\", \"primary_use\", \"year_built\"] but I have a lot of buildings with two or more candidates. Did you figure out how to select the correct one? Download and compare data?\n", "votes": 2 }, { "id": 669901, "postDate": "2019-11-10T16:36:15.507Z", "content": "Yes I compared them manually and merged them to the original test set. Maybe I'll share the kernel later.
", "rawMarkdown": "Yes I compared them manually and merged them to the original test set. Maybe I'll share the kernel later.", "votes": 5 }, { "id": 669909, "postDate": "2019-11-10T16:50:44.033Z", "content": "It's a bit weird that a few test data reading_meter are available. Impact is limited to site 0 currently, did you check if site 0 is in the 22% (with score improvement just by copy/paste the data)? They should be discarded in final score so it will just help us to improve model, otherwise it should be considered as data leakage.
\n\nUpdate: 52 buildings matching on UCF, 7 more to check manually.
\n\n\n[ 0, 2, 3, 4, 6, 7, 8, 10, 11, 16, 18, 20, 22,\n 23, 24, 29, 30, 31, 33, 34, 35, 40, 51, 52, 54, 56,\n 57, 58, 59, 60, 61, 62, 63, 68, 69, 70, 71, 72, 73,\n 74, 75, 76, 77, 78, 86, 87, 88, 91, 94, 97, 98, 103]\n
Update: Also @gunesevitan did you notice the 0.2843 coefficient to match Chilled water?
", "rawMarkdown": "It's a bit weird that a few test data reading_meter are available. Impact is limited to site 0 currently, did you check if site 0 is in the 22% (with score improvement just by copy/paste the data)? They should be discarded in final score so it will just help us to improve model, otherwise it should be considered as data leakage.\n\nUpdate: 52 buildings matching on UCF, 7 more to check manually.\n\n```\n[ 0, 2, 3, 4, 6, 7, 8, 10, 11, 16, 18, 20, 22,\n 23, 24, 29, 30, 31, 33, 34, 35, 40, 51, 52, 54, 56,\n 57, 58, 59, 60, 61, 62, 63, 68, 69, 70, 71, 72, 73,\n 74, 75, 76, 77, 78, 86, 87, 88, 91, 94, 97, 98, 103]\n```\n\nUpdate: Also @gunesevitan did you notice the 0.2843 coefficient to match Chilled water?", "votes": 1 }, { "id": 669982, "postDate": "2019-11-10T20:34:56.693Z", "content": "Did you check the impact on LB of replacing the predictions from your models by the data you scrapped from ucf ? It's a pity if such a leakage indeed exists ..
", "rawMarkdown": "Did you check the impact on LB of replacing the predictions from your models by the data you scrapped from ucf ? It's a pity if such a leakage indeed exists ..", "votes": 1 }, { "id": 670003, "postDate": "2019-11-10T21:56:00.770Z", "content": "I'm going to publish a notebook with this leakage but I'm not using it yet. \nUpdate: Done. https://www.kaggle.com/mpware/ucf-data-limited-leakage
", "rawMarkdown": "I'm going to publish a notebook with this leakage but I'm not using it yet. \nUpdate: Done. https://www.kaggle.com/mpware/ucf-data-limited-leakage", "votes": 4 }, { "id": 670183, "postDate": "2019-11-11T06:27:11.643Z", "content": "@fabiendaniel his team very suddenly and jumped to first place. I'm about 99% sure the scraped data was at least partly responsible. I'm planning to add it myself later this evening, so if I'm suddenly in silver you'll have your answer.
", "rawMarkdown": "@fabiendaniel his team very suddenly and jumped to first place. I'm about 99% sure the scraped data was at least partly responsible. I'm planning to add it myself later this evening, so if I'm suddenly in silver you'll have your answer. " }, { "id": 670197, "postDate": "2019-11-11T06:54:48.907Z", "content": "Since this data leak became public in external data thread several days ago (site 0 = UCF) we tried to check whether parsed data perfectly match those readings in test set, and yes it is:\nhttps://www.kaggle.com/mpware/ucf-data-limited-leakage/comments#670060
", "rawMarkdown": "Since this data leak became public in external data thread several days ago (site 0 = UCF) we tried to check whether parsed data perfectly match those readings in test set, and yes it is:\nhttps://www.kaggle.com/mpware/ucf-data-limited-leakage/comments#670060", "votes": 1 } ] }, { "id": 669788, "postDate": "2019-11-10T13:38:47.507Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 669894, "author_name": "Tolga", "author_url": "", "post_date": "2019-11-10T16:18:31.107000", "content": "They seem to be very rare so they will be washed out by rmsle.
", "votes": 1, "replies": [] }, { "id": 669851, "author_name": "eagle4", "author_url": "", "post_date": "2019-11-10T15:09:27.950000", "content": "Seriously? When are we going to get a clean competition again???
\n\nWho is putting a kernel with all the override that I will upvote for sure?
", "votes": 2, "replies": [] }, { "id": 669766, "author_name": "عثمان", "author_url": "", "post_date": "2019-11-10T12:47:16.050000", "content": "2 submissions ;-)
", "votes": 2, "replies": [] }, { "id": 669688, "author_name": "Mr Loke", "author_url": "", "post_date": "2019-11-10T10:26:26.357000", "content": "I notice the same reading also but it doesn't make sense how can meter reading be in negative value.
", "votes": 2, "replies": [ { "id": 669694, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2019-11-10T10:34:07.900000", "content": "My theory is they are random power outages. That's why meter suddenly outputs a huge negative number (when the electricity goes off). When it comes back, there is a huge positive number.
", "votes": 0, "replies": [] }, { "id": 669737, "author_name": "mezoganet", "author_url": "", "post_date": "2019-11-10T12:13:10.187000", "content": "Or perhaps you just have to change sign ?
\n\nI really don’t know, but this could be an option..
\n\nBut the magnitude is strange anyway. Unless you divide it by 10 ?
", "votes": 0, "replies": [] } ] }, { "id": 681948, "author_name": "Neeraj Kumar", "author_url": "", "post_date": "2019-11-26T18:21:07.517000", "content": "What happens to the test performance when you remove the positive outlier and the negative row?
", "votes": 0, "replies": [] }, { "id": 681851, "author_name": "Fernando Wittmann", "author_url": "", "post_date": "2019-11-26T15:59:05.313000", "content": "The difference between them seems to be the true reading for both readings. For example (1479-1209)/2 = 135 kWh, which is the value of their closest neighbors.
", "votes": 0, "replies": [] }, { "id": 671418, "author_name": "Marcie Wallace", "author_url": "", "post_date": "2019-11-12T17:09:37.950000", "content": "You might have negative readings if power were being produced on site, but that seems unlikely in this case.
", "votes": 0, "replies": [] }, { "id": 669789, "author_name": "MPWARE", "author_url": "", "post_date": "2019-11-10T13:40:20.533000", "content": "Are you talking about this data for building 77? \nhttps://www.oeis.ucf.edu/buildings/68
", "votes": 0, "replies": [ { "id": 669792, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2019-11-10T13:46:12.190000", "content": "Nope, it is from this building.\nhttps://www.oeis.ucf.edu/buildings/131
", "votes": 3, "replies": [] }, { "id": 669895, "author_name": "MPWARE", "author_url": "", "post_date": "2019-11-10T16:28:43.073000", "content": "Thanks, in fact I had 4 candidates for this building:\n
I was able to retrieve and join on [\"square_feet\", \"primary_use\", \"year_built\"] but I have a lot of buildings with two or more candidates. Did you figure out how to select the correct one? Download and compare data?
", "votes": 2, "replies": [] }, { "id": 669901, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2019-11-10T16:36:15.507000", "content": "Yes I compared them manually and merged them to the original test set. Maybe I'll share the kernel later.
", "votes": 5, "replies": [] }, { "id": 669909, "author_name": "MPWARE", "author_url": "", "post_date": "2019-11-10T16:50:44.033000", "content": "It's a bit weird that a few test data reading_meter are available. Impact is limited to site 0 currently, did you check if site 0 is in the 22% (with score improvement just by copy/paste the data)? They should be discarded in final score so it will just help us to improve model, otherwise it should be considered as data leakage.
\n\nUpdate: 52 buildings matching on UCF, 7 more to check manually.
\n\n\n[ 0, 2, 3, 4, 6, 7, 8, 10, 11, 16, 18, 20, 22,\n 23, 24, 29, 30, 31, 33, 34, 35, 40, 51, 52, 54, 56,\n 57, 58, 59, 60, 61, 62, 63, 68, 69, 70, 71, 72, 73,\n 74, 75, 76, 77, 78, 86, 87, 88, 91, 94, 97, 98, 103]\n
Update: Also @gunesevitan did you notice the 0.2843 coefficient to match Chilled water?
", "votes": 1, "replies": [] }, { "id": 669982, "author_name": "FabienDaniel", "author_url": "", "post_date": "2019-11-10T20:34:56.693000", "content": "Did you check the impact on LB of replacing the predictions from your models by the data you scrapped from ucf ? It's a pity if such a leakage indeed exists ..
", "votes": 1, "replies": [] }, { "id": 670003, "author_name": "MPWARE", "author_url": "", "post_date": "2019-11-10T21:56:00.770000", "content": "I'm going to publish a notebook with this leakage but I'm not using it yet. \nUpdate: Done. https://www.kaggle.com/mpware/ucf-data-limited-leakage
", "votes": 4, "replies": [] }, { "id": 670183, "author_name": "Louka Ewington-Pitsos", "author_url": "", "post_date": "2019-11-11T06:27:11.643000", "content": "@fabiendaniel his team very suddenly and jumped to first place. I'm about 99% sure the scraped data was at least partly responsible. I'm planning to add it myself later this evening, so if I'm suddenly in silver you'll have your answer.
", "votes": 0, "replies": [] }, { "id": 670197, "author_name": "Fred Navruzov", "author_url": "", "post_date": "2019-11-11T06:54:48.907000", "content": "Since this data leak became public in external data thread several days ago (site 0 = UCF) we tried to check whether parsed data perfectly match those readings in test set, and yes it is:\nhttps://www.kaggle.com/mpware/ucf-data-limited-leakage/comments#670060
", "votes": 1, "replies": [] } ] }, { "id": 669788, "author_name": "", "author_url": "", "post_date": "2019-11-10T13:38:47.507000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "669524": "I noticed there are some negative readings in site 0 test set. They always occur right before a positive outlier. They don't exist in training set. For example; this is from building_id 77. **(Those are not my predictions, they are from the ucf website)**\n\n\n\n\n\nWhat are they? Are we supposed to predict them or are they going to be excluded from rmsle calculation? ", "669894": "They seem to be very rare so they will be washed out by rmsle.", "669851": "Seriously? When are we going to get a clean competition again???\n\nWho is putting a kernel with all the override that I will upvote for sure?", "669766": "2 submissions ;-)", "669688": "I notice the same reading also but it doesn't make sense how can meter reading be in negative value.", "681948": "What happens to the test performance when you remove the positive outlier and the negative row?", "681851": "The difference between them seems to be the true reading for both readings. For example (1479-1209)/2 = 135 kWh, which is the value of their closest neighbors. ", "671418": "You might have negative readings if power were being produced on site, but that seems unlikely in this case.", "669789": "Are you talking about this data for building 77? \nhttps://www.oeis.ucf.edu/buildings/68\n", "669788": "" } }