First it's nice to realize that I got 1.39 with 2D CNN four months ago and it was my 10-th submission))\nSecond, the key point for the competition - cumsum(X) have 0.99 correlation with Y.
", "messages": [ { "id": "542873", "postDate": "06/04/2019 06:47:41", "content": "First it's nice to realize that I got 1.39 with 2D CNN four months ago and it was my 10-th submission))\nSecond, the key point for the competition - cumsum(X) have 0.99 correlation with Y.
", "rawMarkdown": "First it's nice to realize that I got 1.39 with 2D CNN four months ago and it was my 10-th submission))\nSecond, the key point for the competition - cumsum(X) have 0.99 correlation with Y.", "votes": null }, { "id": "542961", "postDate": "06/04/2019 08:30:45", "content": "I see that CUSUM is a feature of change detection. Well spotted, thanks for the comment!
", "rawMarkdown": "I see that [CUSUM](https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc323.htm) is a feature of [change detection](https://en.wikipedia.org/wiki/Change_detection). Well spotted, thanks for the comment!", "votes": null }, { "id": "542993", "postDate": "06/04/2019 08:57:15", "content": "Hi, @miguel777 , I was following your progress for months, not least because I am from Israel as well. Sorry for you not getting the medal. Can you please explain this 0.99 correlation thing in more detail?
", "rawMarkdown": "Hi, @miguel777 , I was following your progress for months, not least because I am from Israel as well. Sorry for you not getting the medal. Can you please explain this 0.99 correlation thing in more detail?", "votes": null }, { "id": "543038", "postDate": "06/04/2019 09:43:02", "content": "I am sorry, that you missed the medal ranks. Public LB was highly misleading in this comp.
\n\nCould you explain that 0.99 correlation with Y in a bit more detail please? All our features are lightyears aways from such a correlation.
", "rawMarkdown": "I am sorry, that you missed the medal ranks. Public LB was highly misleading in this comp.\n\nCould you explain that 0.99 correlation with Y in a bit more detail please? All our features are lightyears aways from such a correlation.", "votes": null }, { "id": "543063", "postDate": "06/04/2019 10:03:50", "content": "Thanks @zaharch and @ilu000! \nActually, it's an energy based simulation. The cumulative energy of the system have linear dependence with the time. So by simply calculating cumsum(raw data), you can convinced that this measure is inverse proportional to the time. My main feature was cumsum(mel-spectrogram).
", "rawMarkdown": "Thanks @zaharch and @ilu000! \nActually, it's an energy based simulation. The cumulative energy of the system have linear dependence with the time. So by simply calculating cumsum(raw data), you can convinced that this measure is inverse proportional to the time. My main feature was cumsum(mel-spectrogram).", "votes": null }, { "id": "543083", "postDate": "06/04/2019 10:16:46", "content": "I still can't get that to work. So you say that you found a feature that has 0.99 correlation with the train.index? \nIf that extends to test as well, you could easily won this challenge by mapping the TTFs from the academic paper to the test.index. Would you mind creating a quick and dirty kernel out of this?
", "rawMarkdown": "I still can't get that to work. So you say that you found a feature that has 0.99 correlation with the train.index? \nIf that extends to test as well, you could easily won this challenge by mapping the TTFs from the academic paper to the test.index. Would you mind creating a quick and dirty kernel out of this?", "votes": null }, { "id": "543117", "postDate": "06/04/2019 10:39:11", "content": "I also don't understand the cumsum part. If your signal is 150k long then the cumsum is a vector of the same length, why do you call it a feature?
", "rawMarkdown": "I also don't understand the cumsum part. If your signal is 150k long then the cumsum is a vector of the same length, why do you call it a feature?", "votes": null }, { "id": "543120", "postDate": "06/04/2019 10:41:44", "content": "The tricky point here that the train segments are full eatqs [0 17] and test is uniformly distributed on time axis. So, in order to overcome this challenge you need to create a huge amount of cumsum(150k) with all possible t0... But you are right - may be choosing more simple mechanisms then machine learning, you can achieve better results.
", "rawMarkdown": "The tricky point here that the train segments are full eatqs [0 17] and test is uniformly distributed on time axis. So, in order to overcome this challenge you need to create a huge amount of cumsum(150k) with all possible t0... But you are right - may be choosing more simple mechanisms then machine learning, you can achieve better results.", "votes": null } ], "comments": [ { "id": 542961, "author_name": "devilears", "author_url": "", "post_date": "06/04/2019 08:30:45", "content": "I see that CUSUM is a feature of change detection. Well spotted, thanks for the comment!
", "votes": null, "replies": [] }, { "id": 542993, "author_name": "zaharch", "author_url": "", "post_date": "06/04/2019 08:57:15", "content": "Hi, @miguel777 , I was following your progress for months, not least because I am from Israel as well. Sorry for you not getting the medal. Can you please explain this 0.99 correlation thing in more detail?
", "votes": null, "replies": [] }, { "id": 543038, "author_name": "ilu000", "author_url": "", "post_date": "06/04/2019 09:43:02", "content": "I am sorry, that you missed the medal ranks. Public LB was highly misleading in this comp.
\n\nCould you explain that 0.99 correlation with Y in a bit more detail please? All our features are lightyears aways from such a correlation.
", "votes": null, "replies": [] }, { "id": 543063, "author_name": "", "author_url": "", "post_date": "06/04/2019 10:03:50", "content": "Thanks @zaharch and @ilu000! \nActually, it's an energy based simulation. The cumulative energy of the system have linear dependence with the time. So by simply calculating cumsum(raw data), you can convinced that this measure is inverse proportional to the time. My main feature was cumsum(mel-spectrogram).
", "votes": null, "replies": [ { "id": 543083, "author_name": "ilu000", "author_url": "", "post_date": "06/04/2019 10:16:46", "content": "I still can't get that to work. So you say that you found a feature that has 0.99 correlation with the train.index? \nIf that extends to test as well, you could easily won this challenge by mapping the TTFs from the academic paper to the test.index. Would you mind creating a quick and dirty kernel out of this?
", "votes": null, "replies": [] }, { "id": 543117, "author_name": "zaharch", "author_url": "", "post_date": "06/04/2019 10:39:11", "content": "I also don't understand the cumsum part. If your signal is 150k long then the cumsum is a vector of the same length, why do you call it a feature?
", "votes": null, "replies": [] }, { "id": 543120, "author_name": "", "author_url": "", "post_date": "06/04/2019 10:41:44", "content": "The tricky point here that the train segments are full eatqs [0 17] and test is uniformly distributed on time axis. So, in order to overcome this challenge you need to create a huge amount of cumsum(150k) with all possible t0... But you are right - may be choosing more simple mechanisms then machine learning, you can achieve better results.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "542873": "First it's nice to realize that I got 1.39 with 2D CNN four months ago and it was my 10-th submission))\nSecond, the key point for the competition - cumsum(X) have 0.99 correlation with Y.", "542961": "I see that [CUSUM](https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc323.htm) is a feature of [change detection](https://en.wikipedia.org/wiki/Change_detection). Well spotted, thanks for the comment!", "542993": "Hi, @miguel777 , I was following your progress for months, not least because I am from Israel as well. Sorry for you not getting the medal. Can you please explain this 0.99 correlation thing in more detail?", "543038": "I am sorry, that you missed the medal ranks. Public LB was highly misleading in this comp.\n\nCould you explain that 0.99 correlation with Y in a bit more detail please? All our features are lightyears aways from such a correlation.", "543063": "Thanks @zaharch and @ilu000! \nActually, it's an energy based simulation. The cumulative energy of the system have linear dependence with the time. So by simply calculating cumsum(raw data), you can convinced that this measure is inverse proportional to the time. My main feature was cumsum(mel-spectrogram).", "543083": "I still can't get that to work. So you say that you found a feature that has 0.99 correlation with the train.index? \nIf that extends to test as well, you could easily won this challenge by mapping the TTFs from the academic paper to the test.index. Would you mind creating a quick and dirty kernel out of this?", "543117": "I also don't understand the cumsum part. If your signal is 150k long then the cumsum is a vector of the same length, why do you call it a feature?", "543120": "The tricky point here that the train segments are full eatqs [0 17] and test is uniformly distributed on time axis. So, in order to overcome this challenge you need to create a huge amount of cumsum(150k) with all possible t0... But you are right - may be choosing more simple mechanisms then machine learning, you can achieve better results." }, "source": "meta" }