Hi kagglers!
\nPreviously, I proposed the leakage caused by raw timestamps.
\nIn the notebook I published today, I proposed another type of leakage caused by wifi data, the data almost all models are based on.
\nI found the same wifi records are shared among different paths of the training set and even those of the test set. From these shared records, I could partially recover phone user IDs.
I am now realized that the leakage is already widespread.
\nAny comments are welcome!
\nP.S. While I found these leakages and sometimes tried to utilize it after showing it, my own NN models are never better than 11.0 in CV :( . I should read textbooks or take coursera classes, before digging into leakages.
", "messages": [ { "id": "1266628", "postDate": "04/08/2021 00:13:10", "content": "Hi kagglers!
\nPreviously, I proposed the leakage caused by raw timestamps.
\nIn the notebook I published today, I proposed another type of leakage caused by wifi data, the data almost all models are based on.
\nI found the same wifi records are shared among different paths of the training set and even those of the test set. From these shared records, I could partially recover phone user IDs.
I am now realized that the leakage is already widespread.
\nAny comments are welcome!
\nP.S. While I found these leakages and sometimes tried to utilize it after showing it, my own NN models are never better than 11.0 in CV :( . I should read textbooks or take coursera classes, before digging into leakages.
", "rawMarkdown": "Hi kagglers!\n\nPreviously, I [proposed the leakage caused by raw timestamps](https://www.kaggle.com/c/indoor-location-navigation/discussion/228898).\n[In the notebook I published today](https://www.kaggle.com/tomooinubushi/retrieving-user-id-from-leaked-wifi-feature), I proposed another type of leakage caused by wifi data, the data almost all models are based on.\nI found the same wifi records are shared among different paths of the training set and even those of the test set. From these shared records, I could partially recover phone user IDs. \n\nI am now realized that the leakage is already widespread.\n\nAny comments are welcome!\n\nP.S. While I found these leakages and sometimes tried to utilize it after showing it, my own NN models are never better than 11.0 in CV :( . I should read textbooks or take coursera classes, before digging into leakages.", "votes": null }, { "id": "1266719", "postDate": "04/08/2021 02:42:51", "content": "Yes, seems you are completely right. but I think this also means it's very hard to utilize this 'leakage', as everyone already uses this 'leakage' by using wifi features. As you worry, the final score is sadly high compared with the case the train set and the test doesn't share the same users (i.e. no leakage). However, I guess the ranking wouldn't be much different, as everyone already uses this leakage. Of course there may be potential to utilize this 'leakage', though.
", "rawMarkdown": "Yes, seems you are completely right. but I think this also means it's very hard to utilize this 'leakage', as everyone already uses this 'leakage' by using wifi features. As you worry, the final score is sadly high compared with the case the train set and the test doesn't share the same users (i.e. no leakage). However, I guess the ranking wouldn't be much different, as everyone already uses this leakage. Of course there may be potential to utilize this 'leakage', though.", "votes": null }, { "id": "1266754", "postDate": "04/08/2021 04:02:11", "content": "Thank you for your comment.
\nThe influence of the leakage would be not so large, as it is already widespread or just because the leakage has not much strong predictive power. It is consistent with the fact that I could not observe strong changes in the LB ranking this week. You are still 1st, awesome.
I opened the leakage in accordance with the official doc of the kaggle. I thought people should use leakage openly and intentionally. Using leakage without knowing it would make some honest people feel uneasy. I am not always honest as stated above, but I can sympathize with them.
", "rawMarkdown": "Thank you for your comment.\nThe influence of the leakage would be not so large, as it is already widespread or just because the leakage has not much strong predictive power. It is consistent with the fact that I could not observe strong changes in the LB ranking this week. You are still 1st, awesome.\n\nI opened the leakage in accordance with [the official doc of the kaggle](https://www.kaggle.com/docs/competitions#leakage). I thought people should use leakage openly and intentionally. Using leakage without knowing it would make some honest people feel uneasy. I am not always honest as stated above, but I can sympathize with them.", "votes": null } ], "comments": [ { "id": 1266719, "author_name": "mamasinkgs", "author_url": "", "post_date": "04/08/2021 02:42:51", "content": "Yes, seems you are completely right. but I think this also means it's very hard to utilize this 'leakage', as everyone already uses this 'leakage' by using wifi features. As you worry, the final score is sadly high compared with the case the train set and the test doesn't share the same users (i.e. no leakage). However, I guess the ranking wouldn't be much different, as everyone already uses this leakage. Of course there may be potential to utilize this 'leakage', though.
", "votes": null, "replies": [ { "id": 1266754, "author_name": "tomooinubushi", "author_url": "", "post_date": "04/08/2021 04:02:11", "content": "Thank you for your comment.
\nThe influence of the leakage would be not so large, as it is already widespread or just because the leakage has not much strong predictive power. It is consistent with the fact that I could not observe strong changes in the LB ranking this week. You are still 1st, awesome.
I opened the leakage in accordance with the official doc of the kaggle. I thought people should use leakage openly and intentionally. Using leakage without knowing it would make some honest people feel uneasy. I am not always honest as stated above, but I can sympathize with them.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1266628": "Hi kagglers!\n\nPreviously, I [proposed the leakage caused by raw timestamps](https://www.kaggle.com/c/indoor-location-navigation/discussion/228898).\n[In the notebook I published today](https://www.kaggle.com/tomooinubushi/retrieving-user-id-from-leaked-wifi-feature), I proposed another type of leakage caused by wifi data, the data almost all models are based on.\nI found the same wifi records are shared among different paths of the training set and even those of the test set. From these shared records, I could partially recover phone user IDs. \n\nI am now realized that the leakage is already widespread.\n\nAny comments are welcome!\n\nP.S. While I found these leakages and sometimes tried to utilize it after showing it, my own NN models are never better than 11.0 in CV :( . I should read textbooks or take coursera classes, before digging into leakages.", "1266719": "Yes, seems you are completely right. but I think this also means it's very hard to utilize this 'leakage', as everyone already uses this 'leakage' by using wifi features. As you worry, the final score is sadly high compared with the case the train set and the test doesn't share the same users (i.e. no leakage). However, I guess the ranking wouldn't be much different, as everyone already uses this leakage. Of course there may be potential to utilize this 'leakage', though.", "1266754": "Thank you for your comment.\nThe influence of the leakage would be not so large, as it is already widespread or just because the leakage has not much strong predictive power. It is consistent with the fact that I could not observe strong changes in the LB ranking this week. You are still 1st, awesome.\n\nI opened the leakage in accordance with [the official doc of the kaggle](https://www.kaggle.com/docs/competitions#leakage). I thought people should use leakage openly and intentionally. Using leakage without knowing it would make some honest people feel uneasy. I am not always honest as stated above, but I can sympathize with them." }, "source": "meta" }