![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/291388/8691/zipContent.jpeg\")
Given some issues with past competitions, I want to be extra careful and wanted to confirm the train is as expected as this is the first time I see a folder structure like this. \n
Thanks.
", "messages": [ { "id": "291388", "postDate": "03/06/2018 03:15:37", "content": "Given some issues with past competitions, I want to be extra careful and wanted to confirm the train is as expected as this is the first time I see a folder structure like this. \n
Thanks.
", "rawMarkdown": "Given some issues with past competitions, I want to be extra careful and wanted to confirm the train is as expected as this is the first time I see a folder structure like this. \n![enter image description here][1]\n\n\nThanks.\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/291388/8691/zipContent.jpeg", "votes": null }, { "id": "291389", "postDate": "03/06/2018 03:17:56", "content": "Can you re-download and check again? (Regardless, the file should be fine. zip just got excited during the compression.)
", "rawMarkdown": "Can you re-download and check again? (Regardless, the file should be fine. zip just got excited during the compression.)", "votes": null }, { "id": "291393", "postDate": "03/06/2018 03:22:53", "content": "I am lazy to re-download, but I trust you, if this is it, it is.
\n\nFor reference (md5sum):
\n8898d8cfa6a62de21d4e34e4a72a8889 test.csv
\n36da1e8fec8d6765e56d894096a1462e train.csv
Please re-download test.csv. (You're good to go with Train)
\n\nShould be: 8f27a6d1b1f5bcd96c9183654863df98 test.csv
", "rawMarkdown": "Please re-download test.csv. (You're good to go with Train)\n\nShould be: 8f27a6d1b1f5bcd96c9183654863df98 test.csv", "votes": null }, { "id": "291407", "postDate": "03/06/2018 04:08:48", "content": "I cant match the md5sum(test.csv), I just downloaded it again, it still has the full-path structure but I guess that's OK as train matched.
\n\nUnless otherwise notified, I will guess I am just being paranoic here and continue.
", "rawMarkdown": "I cant match the md5sum(test.csv), I just downloaded it again, it still has the full-path structure but I guess that's OK as train matched. \n\nUnless otherwise notified, I will guess I am just being paranoic here and continue.", "votes": null }, { "id": "291479", "postDate": "03/06/2018 07:59:02", "content": "NB The test dataset on the download page is still an old version with 50M samples - until kaggle wake up the only option is to download it from kernels: https://www.kaggle.com/anokas/getting-fixed-test-data
", "rawMarkdown": "NB The test dataset on the download page is still an old version with 50M samples - until kaggle wake up the only option is to download it from kernels: https://www.kaggle.com/anokas/getting-fixed-test-data", "votes": null }, { "id": "291506", "postDate": "03/06/2018 08:54:36", "content": "Thank you for the kind notice.
", "rawMarkdown": "Thank you for the kind notice.", "votes": null }, { "id": "291670", "postDate": "03/06/2018 16:13:48", "content": "Please see this thread:
\n\nhttps://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/51210
", "rawMarkdown": "Please see this thread:\n\nhttps://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/51210", "votes": null }, { "id": "295254", "postDate": "03/13/2018 10:56:35", "content": "Didn't the last talkingdata competition have leakage? Judging from the scores, are we sure this one doesn't too?
", "rawMarkdown": "Didn't the last talkingdata competition have leakage? Judging from the scores, are we sure this one doesn't too?", "votes": null }, { "id": "295258", "postDate": "03/13/2018 11:12:37", "content": "What do you mean by judging from the scores?
", "rawMarkdown": "What do you mean by judging from the scores?", "votes": null }, { "id": "295318", "postDate": "03/13/2018 13:27:38", "content": "Very high AUC
", "rawMarkdown": "Very high AUC", "votes": null }, { "id": "295320", "postDate": "03/13/2018 13:35:56", "content": "Haven't checked in that much detail yet, sorting the scores on IP gives an AUC ~ 0.7 => it does smell funny indeed...
", "rawMarkdown": "Haven't checked in that much detail yet, sorting the scores on IP gives an AUC ~ 0.7 => it does smell funny indeed...", "votes": null }, { "id": "295324", "postDate": "03/13/2018 13:41:36", "content": "I don't think that indicates a leak IMO. It makes logical sense that an IP address (user) who is more likely to download an app in the training set is more likely to download an app a few days later (in the test set) - some people download every app they see while others never download anything.
\n\nSame with apps - certain apps will have higher likelihoods of being downloaded (they may be more appealing etc.) - and this is why you can get 0.9+ just by ordering the apps by their conversion rate in the training set.
", "rawMarkdown": "I don't think that indicates a leak IMO. It makes logical sense that an IP address (user) who is more likely to download an app in the training set is more likely to download an app a few days later (in the test set) - some people download every app they see while others never download anything. \n\nSame with apps - certain apps will have higher likelihoods of being downloaded (they may be more appealing etc.) - and this is why you can get 0.9+ just by ordering the apps by their conversion rate in the training set.", "votes": null }, { "id": "295326", "postDate": "03/13/2018 13:49:50", "content": "In principle I agree, except there have been quite a few competitions where sorting by that sort of thing lead to massive jumps (which under random encoding shouldn't be possible). But I haven't done a rigorous analysis, so it's pure speculation at this point.
", "rawMarkdown": "In principle I agree, except there have been quite a few competitions where sorting by that sort of thing lead to massive jumps (which under random encoding shouldn't be possible). But I haven't done a rigorous analysis, so it's pure speculation at this point.", "votes": null }, { "id": "295328", "postDate": "03/13/2018 13:53:59", "content": "Oh, apologies. I thought you meant sorting the IPs based on their target values in the training set. I agree that just sorting using the random encoding provided shouldn't give >0.5.
\n\nHowever, I don't think the high AUC on it's own is an indicator of a leak. Some problems are just easy to predict.
", "rawMarkdown": "Oh, apologies. I thought you meant sorting the IPs based on their target values in the training set. I agree that just sorting using the random encoding provided shouldn't give >0.5.\n\nHowever, I don't think the high AUC _on it's own_ is an indicator of a leak. Some problems are just easy to predict.", "votes": null }, { "id": "295330", "postDate": "03/13/2018 13:58:31", "content": "No problem - always good to bounce one's thoughts off somebody else.
", "rawMarkdown": "No problem - always good to bounce one's thoughts off somebody else.", "votes": null }, { "id": "295347", "postDate": "03/13/2018 14:29:50", "content": "and there was a similar ID/ sorting leak in the last competition, if I remember correctly
", "rawMarkdown": "and there was a similar ID/ sorting leak in the last competition, if I remember correctly", "votes": null }, { "id": "295348", "postDate": "03/13/2018 14:31:07", "content": "and if I was the company and could easily get 96 AUC I wouldn't pay $$ to outsource the problem
", "rawMarkdown": "and if I was the company and could easily get 96 AUC I wouldn't pay $$ to outsource the problem", "votes": null }, { "id": "295356", "postDate": "03/13/2018 14:50:27", "content": "Well, that's a more general phenomenon, case in point: the Toxic Comments contest. Given a NB-SVM benchmark scoring ~ .982 AUC while running on a laptop, would any business care about an advanced deep learning model achieving .988 on a cluster?
", "rawMarkdown": "Well, that's a more general phenomenon, case in point: the Toxic Comments contest. Given a NB-SVM benchmark scoring ~ .982 AUC while running on a laptop, would any business care about an advanced deep learning model achieving .988 on a cluster?", "votes": null } ], "comments": [ { "id": 291389, "author_name": "inversion", "author_url": "", "post_date": "03/06/2018 03:17:56", "content": "Can you re-download and check again? (Regardless, the file should be fine. zip just got excited during the compression.)
", "votes": null, "replies": [ { "id": 291393, "author_name": "carloshuertas", "author_url": "", "post_date": "03/06/2018 03:22:53", "content": "I am lazy to re-download, but I trust you, if this is it, it is.
\n\nFor reference (md5sum):
\n8898d8cfa6a62de21d4e34e4a72a8889 test.csv
\n36da1e8fec8d6765e56d894096a1462e train.csv
Please re-download test.csv. (You're good to go with Train)
\n\nShould be: 8f27a6d1b1f5bcd96c9183654863df98 test.csv
", "votes": null, "replies": [] }, { "id": 291407, "author_name": "carloshuertas", "author_url": "", "post_date": "03/06/2018 04:08:48", "content": "I cant match the md5sum(test.csv), I just downloaded it again, it still has the full-path structure but I guess that's OK as train matched.
\n\nUnless otherwise notified, I will guess I am just being paranoic here and continue.
", "votes": null, "replies": [] }, { "id": 291670, "author_name": "inversion", "author_url": "", "post_date": "03/06/2018 16:13:48", "content": "Please see this thread:
\n\nhttps://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/51210
", "votes": null, "replies": [] } ] }, { "id": 291479, "author_name": "anokas", "author_url": "", "post_date": "03/06/2018 07:59:02", "content": "NB The test dataset on the download page is still an old version with 50M samples - until kaggle wake up the only option is to download it from kernels: https://www.kaggle.com/anokas/getting-fixed-test-data
", "votes": null, "replies": [ { "id": 291506, "author_name": "wenpengwei", "author_url": "", "post_date": "03/06/2018 08:54:36", "content": "Thank you for the kind notice.
", "votes": null, "replies": [] } ] }, { "id": 295254, "author_name": "domcastro", "author_url": "", "post_date": "03/13/2018 10:56:35", "content": "Didn't the last talkingdata competition have leakage? Judging from the scores, are we sure this one doesn't too?
", "votes": null, "replies": [ { "id": 295258, "author_name": "anokas", "author_url": "", "post_date": "03/13/2018 11:12:37", "content": "What do you mean by judging from the scores?
", "votes": null, "replies": [] }, { "id": 295318, "author_name": "domcastro", "author_url": "", "post_date": "03/13/2018 13:27:38", "content": "Very high AUC
", "votes": null, "replies": [] }, { "id": 295320, "author_name": "konradb", "author_url": "", "post_date": "03/13/2018 13:35:56", "content": "Haven't checked in that much detail yet, sorting the scores on IP gives an AUC ~ 0.7 => it does smell funny indeed...
", "votes": null, "replies": [] }, { "id": 295324, "author_name": "anokas", "author_url": "", "post_date": "03/13/2018 13:41:36", "content": "I don't think that indicates a leak IMO. It makes logical sense that an IP address (user) who is more likely to download an app in the training set is more likely to download an app a few days later (in the test set) - some people download every app they see while others never download anything.
\n\nSame with apps - certain apps will have higher likelihoods of being downloaded (they may be more appealing etc.) - and this is why you can get 0.9+ just by ordering the apps by their conversion rate in the training set.
", "votes": null, "replies": [] }, { "id": 295326, "author_name": "konradb", "author_url": "", "post_date": "03/13/2018 13:49:50", "content": "In principle I agree, except there have been quite a few competitions where sorting by that sort of thing lead to massive jumps (which under random encoding shouldn't be possible). But I haven't done a rigorous analysis, so it's pure speculation at this point.
", "votes": null, "replies": [] }, { "id": 295328, "author_name": "anokas", "author_url": "", "post_date": "03/13/2018 13:53:59", "content": "Oh, apologies. I thought you meant sorting the IPs based on their target values in the training set. I agree that just sorting using the random encoding provided shouldn't give >0.5.
\n\nHowever, I don't think the high AUC on it's own is an indicator of a leak. Some problems are just easy to predict.
", "votes": null, "replies": [] }, { "id": 295330, "author_name": "konradb", "author_url": "", "post_date": "03/13/2018 13:58:31", "content": "No problem - always good to bounce one's thoughts off somebody else.
", "votes": null, "replies": [] }, { "id": 295347, "author_name": "domcastro", "author_url": "", "post_date": "03/13/2018 14:29:50", "content": "and there was a similar ID/ sorting leak in the last competition, if I remember correctly
", "votes": null, "replies": [] }, { "id": 295348, "author_name": "domcastro", "author_url": "", "post_date": "03/13/2018 14:31:07", "content": "and if I was the company and could easily get 96 AUC I wouldn't pay $$ to outsource the problem
", "votes": null, "replies": [] }, { "id": 295356, "author_name": "konradb", "author_url": "", "post_date": "03/13/2018 14:50:27", "content": "Well, that's a more general phenomenon, case in point: the Toxic Comments contest. Given a NB-SVM benchmark scoring ~ .982 AUC while running on a laptop, would any business care about an advanced deep learning model achieving .988 on a cluster?
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "291388": "Given some issues with past competitions, I want to be extra careful and wanted to confirm the train is as expected as this is the first time I see a folder structure like this. \n![enter image description here][1]\n\n\nThanks.\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/291388/8691/zipContent.jpeg", "291389": "Can you re-download and check again? (Regardless, the file should be fine. zip just got excited during the compression.)", "291393": "I am lazy to re-download, but I trust you, if this is it, it is.\n\nFor reference (md5sum): \n8898d8cfa6a62de21d4e34e4a72a8889 test.csv \n36da1e8fec8d6765e56d894096a1462e train.csv", "291398": "Please re-download test.csv. (You're good to go with Train)\n\nShould be: 8f27a6d1b1f5bcd96c9183654863df98 test.csv", "291407": "I cant match the md5sum(test.csv), I just downloaded it again, it still has the full-path structure but I guess that's OK as train matched. \n\nUnless otherwise notified, I will guess I am just being paranoic here and continue.", "291479": "NB The test dataset on the download page is still an old version with 50M samples - until kaggle wake up the only option is to download it from kernels: https://www.kaggle.com/anokas/getting-fixed-test-data", "291506": "Thank you for the kind notice.", "291670": "Please see this thread:\n\nhttps://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/51210", "295254": "Didn't the last talkingdata competition have leakage? Judging from the scores, are we sure this one doesn't too?", "295258": "What do you mean by judging from the scores?", "295318": "Very high AUC", "295320": "Haven't checked in that much detail yet, sorting the scores on IP gives an AUC ~ 0.7 => it does smell funny indeed...", "295324": "I don't think that indicates a leak IMO. It makes logical sense that an IP address (user) who is more likely to download an app in the training set is more likely to download an app a few days later (in the test set) - some people download every app they see while others never download anything. \n\nSame with apps - certain apps will have higher likelihoods of being downloaded (they may be more appealing etc.) - and this is why you can get 0.9+ just by ordering the apps by their conversion rate in the training set.", "295326": "In principle I agree, except there have been quite a few competitions where sorting by that sort of thing lead to massive jumps (which under random encoding shouldn't be possible). But I haven't done a rigorous analysis, so it's pure speculation at this point.", "295328": "Oh, apologies. I thought you meant sorting the IPs based on their target values in the training set. I agree that just sorting using the random encoding provided shouldn't give >0.5.\n\nHowever, I don't think the high AUC _on it's own_ is an indicator of a leak. Some problems are just easy to predict.", "295330": "No problem - always good to bounce one's thoughts off somebody else.", "295347": "and there was a similar ID/ sorting leak in the last competition, if I remember correctly", "295348": "and if I was the company and could easily get 96 AUC I wouldn't pay $$ to outsource the problem", "295356": "Well, that's a more general phenomenon, case in point: the Toxic Comments contest. Given a NB-SVM benchmark scoring ~ .982 AUC while running on a laptop, would any business care about an advanced deep learning model achieving .988 on a cluster?" }, "source": "meta" }