If it was how do you guarantee that it is directly benefitted from the leaks. I don't know as I did not try and will not, and I don't see anybody who made it apparently.
", "messages": [ { "id": "680020", "postDate": "11/23/2019 20:44:28", "content": "If it was how do you guarantee that it is directly benefitted from the leaks. I don't know as I did not try and will not, and I don't see anybody who made it apparently.
", "rawMarkdown": "If it was how do you guarantee that it is directly benefitted from the leaks. I don't know as I did not try and will not, and I don't see anybody who made it apparently.", "votes": null }, { "id": "680028", "postDate": "11/23/2019 21:31:53", "content": "I made an attempt at this using site0 data to try to improve score for sites 1-15, but I saw mixed results. I think you're asking the right question though.
", "rawMarkdown": "I made an attempt at this using site0 data to try to improve score for sites 1-15, but I saw mixed results. I think you're asking the right question though.", "votes": null }, { "id": "680991", "postDate": "11/25/2019 13:47:56", "content": "What is the mixed results? So, There is nobody(but you) willing to say about this topic. XD I don't know if someone has significant advantages from the leaks(except replacing) when I see LB histories seemingly.
", "rawMarkdown": "What is the mixed results? So, There is nobody(but you) willing to say about this topic. XD I don't know if someone has significant advantages from the leaks(except replacing) when I see LB histories seemingly.", "votes": null }, { "id": "681015", "postDate": "11/25/2019 14:35:37", "content": "i tried with site 1 and 0 but lb score decreased by 0.01 (no idea why )
", "rawMarkdown": "i tried with site 1 and 0 but lb score decreased by 0.01 (no idea why )", "votes": null }, { "id": "681126", "postDate": "11/25/2019 17:47:06", "content": "So it was worse? This might have to be asked by ourselves. How does it guarantee that data from the leaks are significantly helpful to predict the other sites? just nothing but augmenting some proportion of train dataset, so potentially helpful by the law of large numbers or nothing different from before leaks????
\n\nWe know one of the worst scenarios. Someone hide publicly available leaks and blend some of it at the end of competition. Who will be going to do and explain their secret when they get gold or above? looks fun. silver/bronze might be… I cannot see how LB will be changed, it is so contaminated, and I cannot aim it not too low position.
", "rawMarkdown": "So it was worse? This might have to be asked by ourselves. How does it guarantee that data from the leaks are significantly helpful to predict the other sites? just nothing but augmenting some proportion of train dataset, so potentially helpful by the law of large numbers or nothing different from before leaks????\n\nWe know one of the worst scenarios. Someone hide publicly available leaks and blend some of it at the end of competition. Who will be going to do and explain their secret when they get gold or above? looks fun. silver/bronze might be… I cannot see how LB will be changed, it is so contaminated, and I cannot aim it not too low position.", "votes": null }, { "id": "681215", "postDate": "11/25/2019 21:02:56", "content": "Hi @hanjoonchoe I was able to increase my score by 0.01 by using leak data as additional training data. To eliminate the effect of the train data on the test data I compared 2 submission where the leak data was used to update the test portion for that site.\n1 submission I had not used the leak data to train...the other submission I did use. It gave a nice performance boost.
\n\nI've only used 1 site with leak data sofar but will be adding more in a similar way.
\n\nIt could be coincidence or the way I put it together...for me at least it helped.
", "rawMarkdown": "Hi @hanjoonchoe I was able to increase my score by 0.01 by using leak data as additional training data. To eliminate the effect of the train data on the test data I compared 2 submission where the leak data was used to update the test portion for that site.\n1 submission I had not used the leak data to train...the other submission I did use. It gave a nice performance boost.\n\nI've only used 1 site with leak data sofar but will be adding more in a similar way.\n\nIt could be coincidence or the way I put it together...for me at least it helped.", "votes": null }, { "id": "681220", "postDate": "11/25/2019 21:16:37", "content": "Thank you for notice. ;-)
", "rawMarkdown": "Thank you for notice. ;-)", "votes": null } ], "comments": [ { "id": 680028, "author_name": "teeyee314", "author_url": "", "post_date": "11/23/2019 21:31:53", "content": "I made an attempt at this using site0 data to try to improve score for sites 1-15, but I saw mixed results. I think you're asking the right question though.
", "votes": null, "replies": [ { "id": 680991, "author_name": "hanjoonchoe", "author_url": "", "post_date": "11/25/2019 13:47:56", "content": "What is the mixed results? So, There is nobody(but you) willing to say about this topic. XD I don't know if someone has significant advantages from the leaks(except replacing) when I see LB histories seemingly.
", "votes": null, "replies": [] } ] }, { "id": 681015, "author_name": "negi009", "author_url": "", "post_date": "11/25/2019 14:35:37", "content": "i tried with site 1 and 0 but lb score decreased by 0.01 (no idea why )
", "votes": null, "replies": [ { "id": 681126, "author_name": "hanjoonchoe", "author_url": "", "post_date": "11/25/2019 17:47:06", "content": "So it was worse? This might have to be asked by ourselves. How does it guarantee that data from the leaks are significantly helpful to predict the other sites? just nothing but augmenting some proportion of train dataset, so potentially helpful by the law of large numbers or nothing different from before leaks????
\n\nWe know one of the worst scenarios. Someone hide publicly available leaks and blend some of it at the end of competition. Who will be going to do and explain their secret when they get gold or above? looks fun. silver/bronze might be… I cannot see how LB will be changed, it is so contaminated, and I cannot aim it not too low position.
", "votes": null, "replies": [] } ] }, { "id": 681215, "author_name": "rsmits", "author_url": "", "post_date": "11/25/2019 21:02:56", "content": "Hi @hanjoonchoe I was able to increase my score by 0.01 by using leak data as additional training data. To eliminate the effect of the train data on the test data I compared 2 submission where the leak data was used to update the test portion for that site.\n1 submission I had not used the leak data to train...the other submission I did use. It gave a nice performance boost.
\n\nI've only used 1 site with leak data sofar but will be adding more in a similar way.
\n\nIt could be coincidence or the way I put it together...for me at least it helped.
", "votes": null, "replies": [ { "id": 681220, "author_name": "hanjoonchoe", "author_url": "", "post_date": "11/25/2019 21:16:37", "content": "Thank you for notice. ;-)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "680020": "If it was how do you guarantee that it is directly benefitted from the leaks. I don't know as I did not try and will not, and I don't see anybody who made it apparently.", "680028": "I made an attempt at this using site0 data to try to improve score for sites 1-15, but I saw mixed results. I think you're asking the right question though.", "680991": "What is the mixed results? So, There is nobody(but you) willing to say about this topic. XD I don't know if someone has significant advantages from the leaks(except replacing) when I see LB histories seemingly.", "681015": "i tried with site 1 and 0 but lb score decreased by 0.01 (no idea why )", "681126": "So it was worse? This might have to be asked by ourselves. How does it guarantee that data from the leaks are significantly helpful to predict the other sites? just nothing but augmenting some proportion of train dataset, so potentially helpful by the law of large numbers or nothing different from before leaks????\n\nWe know one of the worst scenarios. Someone hide publicly available leaks and blend some of it at the end of competition. Who will be going to do and explain their secret when they get gold or above? looks fun. silver/bronze might be… I cannot see how LB will be changed, it is so contaminated, and I cannot aim it not too low position.", "681215": "Hi @hanjoonchoe I was able to increase my score by 0.01 by using leak data as additional training data. To eliminate the effect of the train data on the test data I compared 2 submission where the leak data was used to update the test portion for that site.\n1 submission I had not used the leak data to train...the other submission I did use. It gave a nice performance boost.\n\nI've only used 1 site with leak data sofar but will be adding more in a similar way.\n\nIt could be coincidence or the way I put it together...for me at least it helped.", "681220": "Thank you for notice. ;-)" }, "source": "meta" }