As described in instructions, Isn't use of scaler= StandardScaler() for test data invalid and will disqualifies the modal? \nSimilarly filtering complete signal at once, is it allowed. As most of scripts scoring 0.8+ are using it and getting score in leader board
\n\nCan admin help to understand this?
\n\nWhen predicting, you may not use data from the future! In other words, if you are predicting labels for id subj1_series9_11, you may not incorporate any frame after 11 of series 9 from subject 1.
\n\nYou must be careful not to include data leakage. For example, you may not center the signals for subj1_series9_11 based on the mean of all frames for subj1_series9. Instead, you must use the mean based on frame 0 to 10.
\n\nAs described in instructions, Isn't use of scaler= StandardScaler() for test data invalid and will disqualifies the modal? \nSimilarly filtering complete signal at once, is it allowed. As most of scripts scoring 0.8+ are using it and getting score in leader board
\n\nCan admin help to understand this?
\n\nWhen predicting, you may not use data from the future! In other words, if you are predicting labels for id subj1_series9_11, you may not incorporate any frame after 11 of series 9 from subject 1.
\n\nYou must be careful not to include data leakage. For example, you may not center the signals for subj1_series9_11 based on the mean of all frames for subj1_series9. Instead, you must use the mean based on frame 0 to 10.
\n\nThe way I understand it is your classifier must be CAUSAL when classifying series 9 and 10.
\n\nIt's ok to train your algorithm on all the training data but when predicting series 9 and 10 you can't use data from time k+1,... to classify the events at time k.
\n\nAs such you can apply a causal filter (no filtfilt) to your whole signal as long as when using this filtered signal for classification, your predictor for the output at time k only uses the filtered signal from times 1 to k.
\n\nSo basically your predictor for y(k) = function(x(1),...,x(k))
", "rawMarkdown": "The way I understand it is your classifier must be CAUSAL when classifying series 9 and 10.\r\n\r\nIt's ok to train your algorithm on all the training data but when predicting series 9 and 10 you can't use data from time k+1,... to classify the events at time k.\r\n\r\nAs such you can apply a causal filter (no filtfilt) to your whole signal as long as when using this filtered signal for classification, your predictor for the output at time k only uses the filtered signal from times 1 to k.\r\n\r\nSo basically your predictor for y(k) = function(x(1),...,x(k))", "votes": null }, { "id": "86952", "postDate": "07/25/2015 15:47:54", "content": "yes, filter should be causal and as butterworth filter is used its great, what about scaler= StandardScaler() this normalize the signal by mean and variance.
", "rawMarkdown": "yes, filter should be causal and as butterworth filter is used its great, what about *scaler= StandardScaler()* this normalize the signal by mean and variance.", "votes": null }, { "id": "86967", "postDate": "07/25/2015 20:28:38", "content": "Scaling the whole dataset by its mean and variance is obviously not OK for series 9 and 10 since the scaled data will be a function of the whole dataset.
\n\nThat being said, I'd say it's OK to do it for fitting the model on series 1-8 and then using the same scaling values for series 9-10.
", "rawMarkdown": "Scaling the whole dataset by its mean and variance is obviously not OK for series 9 and 10 since the scaled data will be a function of the whole dataset.\r\n\r\nThat being said, I'd say it's OK to do it for fitting the model on series 1-8 and then using the same scaling values for series 9-10.", "votes": null }, { "id": "86970", "postDate": "07/25/2015 20:50:42", "content": "As far as I can see a butterworth filter would also violate the rules, no? While it does not have long tails in the time domain it should still have some leakage into the future, or am I missing something?
", "rawMarkdown": "As far as I can see a butterworth filter would also violate the rules, no? While it does not have long tails in the time domain it should still have some leakage into the future, or am I missing something?", "votes": null }, { "id": "86979", "postDate": "07/25/2015 23:11:00", "content": "It will depend on your language of choice but for instance in matlab whenever you use the filter function, no matter what coefficients you pass it (for a butterworth, chebychev, elliptical, FIR,...) you'll always get a causal filter.
\n\nIt basically depends how the filter is implemented. If you take a FIR filter b(q^-1) and "correct" the lag by b(q^-1)*q^(N/2), you get a noncausal filter.
\n\nIf you're in doubt how your software is implementing the filter try computing its inpulse response. You can pass it a kronecker delta u=[0,...,0,1,0,...] and if the filtered response is zero before the index of the 1 your filter is by definition causal. I.e., you should get y=[0,..,0,h(0),h(1),...]
", "rawMarkdown": "It will depend on your language of choice but for instance in matlab whenever you use the filter function, no matter what coefficients you pass it (for a butterworth, chebychev, elliptical, FIR,...) you'll always get a causal filter.\r\n\r\nIt basically depends how the filter is implemented. If you take a FIR filter b(q^-1) and \"correct\" the lag by b(q^-1)*q^(N/2), you get a noncausal filter.\r\n\r\nIf you're in doubt how your software is implementing the filter try computing its inpulse response. You can pass it a kronecker delta u=[0,...,0,1,0,...] and if the filtered response is zero before the index of the 1 your filter is by definition causal. I.e., you should get y=[0,..,0,h(0),h(1),...]", "votes": null }, { "id": "88354", "postDate": "08/04/2015 12:59:48", "content": "If you are using python, using the lfilter from scipy is okay with the competition rules.
\n\nhttp://dsp.stackexchange.com/questions/19084/applying-filter-in-scipy-signal-use-lfilter-or-filtfilt
", "rawMarkdown": "If you are using python, using the lfilter from scipy is okay with the competition rules.\r\n\r\nhttp://dsp.stackexchange.com/questions/19084/applying-filter-in-scipy-signal-use-lfilter-or-filtfilt", "votes": null } ], "comments": [ { "id": 86951, "author_name": "ajoo88", "author_url": "", "post_date": "07/25/2015 15:10:37", "content": "The way I understand it is your classifier must be CAUSAL when classifying series 9 and 10.
\n\nIt's ok to train your algorithm on all the training data but when predicting series 9 and 10 you can't use data from time k+1,... to classify the events at time k.
\n\nAs such you can apply a causal filter (no filtfilt) to your whole signal as long as when using this filtered signal for classification, your predictor for the output at time k only uses the filtered signal from times 1 to k.
\n\nSo basically your predictor for y(k) = function(x(1),...,x(k))
", "votes": null, "replies": [] }, { "id": 86952, "author_name": "nikeshbajaj", "author_url": "", "post_date": "07/25/2015 15:47:54", "content": "yes, filter should be causal and as butterworth filter is used its great, what about scaler= StandardScaler() this normalize the signal by mean and variance.
", "votes": null, "replies": [] }, { "id": 86967, "author_name": "ajoo88", "author_url": "", "post_date": "07/25/2015 20:28:38", "content": "Scaling the whole dataset by its mean and variance is obviously not OK for series 9 and 10 since the scaled data will be a function of the whole dataset.
\n\nThat being said, I'd say it's OK to do it for fitting the model on series 1-8 and then using the same scaling values for series 9-10.
", "votes": null, "replies": [] }, { "id": 86970, "author_name": "hasselmann", "author_url": "", "post_date": "07/25/2015 20:50:42", "content": "As far as I can see a butterworth filter would also violate the rules, no? While it does not have long tails in the time domain it should still have some leakage into the future, or am I missing something?
", "votes": null, "replies": [] }, { "id": 86979, "author_name": "ajoo88", "author_url": "", "post_date": "07/25/2015 23:11:00", "content": "It will depend on your language of choice but for instance in matlab whenever you use the filter function, no matter what coefficients you pass it (for a butterworth, chebychev, elliptical, FIR,...) you'll always get a causal filter.
\n\nIt basically depends how the filter is implemented. If you take a FIR filter b(q^-1) and "correct" the lag by b(q^-1)*q^(N/2), you get a noncausal filter.
\n\nIf you're in doubt how your software is implementing the filter try computing its inpulse response. You can pass it a kronecker delta u=[0,...,0,1,0,...] and if the filtered response is zero before the index of the 1 your filter is by definition causal. I.e., you should get y=[0,..,0,h(0),h(1),...]
", "votes": null, "replies": [] }, { "id": 88354, "author_name": "mrbeer", "author_url": "", "post_date": "08/04/2015 12:59:48", "content": "If you are using python, using the lfilter from scipy is okay with the competition rules.
\n\nhttp://dsp.stackexchange.com/questions/19084/applying-filter-in-scipy-signal-use-lfilter-or-filtfilt
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "86916": "As described in instructions, Isn't use of *scaler= StandardScaler()* for test data **invalid and will disqualifies the modal?** \r\nSimilarly filtering complete signal at once, is it allowed. As most of scripts scoring 0.8+ are using it and getting score in leader board\r\n\r\nCan admin help to understand this?\r\n___ \r\n**When predicting, you may not use data from the future!** In other words, if you are predicting labels for id subj1_series9_11, you may not incorporate any frame after 11 of series 9 from subject 1.\r\n\r\n**You must be careful not to include data leakage.** For example, you may not center the signals for subj1_series9_11 based on the mean of all frames for subj1_series9. Instead, you must use the mean based on frame 0 to 10.\r\n___", "86951": "The way I understand it is your classifier must be CAUSAL when classifying series 9 and 10.\r\n\r\nIt's ok to train your algorithm on all the training data but when predicting series 9 and 10 you can't use data from time k+1,... to classify the events at time k.\r\n\r\nAs such you can apply a causal filter (no filtfilt) to your whole signal as long as when using this filtered signal for classification, your predictor for the output at time k only uses the filtered signal from times 1 to k.\r\n\r\nSo basically your predictor for y(k) = function(x(1),...,x(k))", "86952": "yes, filter should be causal and as butterworth filter is used its great, what about *scaler= StandardScaler()* this normalize the signal by mean and variance.", "86967": "Scaling the whole dataset by its mean and variance is obviously not OK for series 9 and 10 since the scaled data will be a function of the whole dataset.\r\n\r\nThat being said, I'd say it's OK to do it for fitting the model on series 1-8 and then using the same scaling values for series 9-10.", "86970": "As far as I can see a butterworth filter would also violate the rules, no? While it does not have long tails in the time domain it should still have some leakage into the future, or am I missing something?", "86979": "It will depend on your language of choice but for instance in matlab whenever you use the filter function, no matter what coefficients you pass it (for a butterworth, chebychev, elliptical, FIR,...) you'll always get a causal filter.\r\n\r\nIt basically depends how the filter is implemented. If you take a FIR filter b(q^-1) and \"correct\" the lag by b(q^-1)*q^(N/2), you get a noncausal filter.\r\n\r\nIf you're in doubt how your software is implementing the filter try computing its inpulse response. You can pass it a kronecker delta u=[0,...,0,1,0,...] and if the filtered response is zero before the index of the 1 your filter is by definition causal. I.e., you should get y=[0,..,0,h(0),h(1),...]", "88354": "If you are using python, using the lfilter from scipy is okay with the competition rules.\r\n\r\nhttp://dsp.stackexchange.com/questions/19084/applying-filter-in-scipy-signal-use-lfilter-or-filtfilt" }, "source": "meta" }