Since this is a feature building competition, it would be a good idea to know how many features is everyone using :)
\n\nApprox. 60 features / LB 1.505
", "messages": [ { "id": 455565, "postDate": "2019-01-14T07:04:41.210Z", "content": "Since this is a feature building competition, it would be a good idea to know how many features is everyone using :)
\n\nApprox. 60 features / LB 1.505
", "rawMarkdown": "Since this is a feature building competition, it would be a good idea to know how many features is everyone using :)\n\n\nApprox. 60 features / LB 1.505\n\n", "votes": 10 }, { "id": 456160, "postDate": "2019-01-15T08:52:10.460Z", "content": "currently, I am using just few aggregated features, as following:\n- mean, std, var, min, max, abs_max / segment;\n- A0-A10: being the FFT module values for CC and first 10 harmonics; R1-R5: being the real components of 1st 5 harmonics;\nTotally: 21 features / LB: 1.884 (very poor).
\n\nI am also calculated 10 times more segments than the 4194 resulting by simply dividing the train data in 150k segments (I simply shift segments with a shift = 1/10 * segment, a technique for data multiplication we used to use in the past: we call it 'shifting aperture').
\n\nIt looks that my validation error is quite close to the LB score, but the train error is much smaller so I do have a major problem with my model, a lot to improve to it.
", "rawMarkdown": "currently, I am using just few aggregated features, as following:\n- mean, std, var, min, max, abs_max / segment;\n- A0-A10: being the FFT module values for CC and first 10 harmonics; R1-R5: being the real components of 1st 5 harmonics;\nTotally: 21 features / LB: 1.884 (very poor).\n\nI am also calculated 10 times more segments than the 4194 resulting by simply dividing the train data in 150k segments (I simply shift segments with a shift = 1/10 * segment, a technique for data multiplication we used to use in the past: we call it 'shifting aperture').\n\nIt looks that my validation error is quite close to the LB score, but the train error is much smaller so I do have a major problem with my model, a lot to improve to it.", "votes": 5, "replies": [ { "id": 456329, "postDate": "2019-01-15T15:26:47.050Z", "content": "Seems like you may not be treating you train / validation sets in the same way, or your validation set does not resemble the test set.
", "rawMarkdown": "Seems like you may not be treating you train / validation sets in the same way, or your validation set does not resemble the test set." }, { "id": 458202, "postDate": "2019-01-19T03:59:42.723Z", "content": "Hi! Gabriel, I been using shifting aperture, but the more shifts I add my LB Score gets worst, I'm trying to find the sweet spot to calibrate my CV with LB based on the number of shifts
", "rawMarkdown": "Hi! Gabriel, I been using shifting aperture, but the more shifts I add my LB Score gets worst, I'm trying to find the sweet spot to calibrate my CV with LB based on the number of shifts" }, { "id": 460285, "postDate": "2019-01-23T11:11:54.713Z", "content": "I used welch using a point shift (np.arange(100,1000,100))and got 1.62 using Genetic Programming and 1.67 with lgb - might be useful for an ensemble
", "rawMarkdown": "I used welch using a point shift (np.arange(100,1000,100))and got 1.62 using Genetic Programming and 1.67 with lgb - might be useful for an ensemble" } ] }, { "id": 460784, "postDate": "2019-01-24T12:17:15.017Z", "content": "Andrew's Script + GP == 133 features for a score of 1.461
", "rawMarkdown": "Andrew's Script + GP == 133 features for a score of 1.461", "votes": 4, "replies": [ { "id": 460796, "postDate": "2019-01-24T12:46:41.447Z", "content": "gaussian process?
", "rawMarkdown": "gaussian process?", "votes": 1 }, { "id": 460797, "postDate": "2019-01-24T12:48:52.273Z", "content": "Apologies, GP = Genetic Programming
", "rawMarkdown": "Apologies, GP = Genetic Programming", "votes": 1 }, { "id": 460800, "postDate": "2019-01-24T12:52:17.163Z", "content": "saw the previous post later. good stuff!
", "rawMarkdown": "saw the previous post later. good stuff!", "votes": 1 }, { "id": 460802, "postDate": "2019-01-24T13:03:40.190Z", "content": "@Abhishek, GP is so 2nd nature to @Scirpus he thinks we all hear Genetic Programming when we see GP instead of gaussian process or general practitioner :-)
", "rawMarkdown": "@Abhishek, GP is so 2nd nature to @Scirpus he thinks we all hear Genetic Programming when we see GP instead of gaussian process or general practitioner :-)", "votes": 2 }, { "id": 460805, "postDate": "2019-01-24T13:07:05.463Z", "content": "LOL so true ;)
", "rawMarkdown": "LOL so true ;)", "votes": 2 }, { "id": 509019, "postDate": "2019-04-07T07:56:59.527Z", "content": "what is Andy's script - sorry for the newb questoin
", "rawMarkdown": "what is Andy's script - sorry for the newb questoin" }, { "id": 509292, "postDate": "2019-04-07T16:26:32.577Z", "content": "I expect that they mean the rather popular kernel by Andrew Lukyanenko , see https://www.kaggle.com/artgor/earthquakes-fe-more-features-and-samples
", "rawMarkdown": "I expect that they mean the rather popular kernel by Andrew Lukyanenko , see [https://www.kaggle.com/artgor/earthquakes-fe-more-features-and-samples](https://www.kaggle.com/artgor/earthquakes-fe-more-features-and-samples)" } ] }, { "id": 455698, "postDate": "2019-01-14T11:42:24.620Z", "rawMarkdown": "", "isDeleted": true }, { "id": 464988, "postDate": "2019-02-02T00:52:50.840Z", "content": "Thanks
", "rawMarkdown": "Thanks" } ], "comments": [ { "id": 456160, "author_name": "Gabriel Preda", "author_url": "", "post_date": "2019-01-15T08:52:10.460000", "content": "currently, I am using just few aggregated features, as following:\n- mean, std, var, min, max, abs_max / segment;\n- A0-A10: being the FFT module values for CC and first 10 harmonics; R1-R5: being the real components of 1st 5 harmonics;\nTotally: 21 features / LB: 1.884 (very poor).
\n\nI am also calculated 10 times more segments than the 4194 resulting by simply dividing the train data in 150k segments (I simply shift segments with a shift = 1/10 * segment, a technique for data multiplication we used to use in the past: we call it 'shifting aperture').
\n\nIt looks that my validation error is quite close to the LB score, but the train error is much smaller so I do have a major problem with my model, a lot to improve to it.
", "votes": 5, "replies": [ { "id": 456329, "author_name": "HorizonPicking2k18", "author_url": "", "post_date": "2019-01-15T15:26:47.050000", "content": "Seems like you may not be treating you train / validation sets in the same way, or your validation set does not resemble the test set.
", "votes": 0, "replies": [] }, { "id": 458202, "author_name": "C4rl05/V", "author_url": "", "post_date": "2019-01-19T03:59:42.723000", "content": "Hi! Gabriel, I been using shifting aperture, but the more shifts I add my LB Score gets worst, I'm trying to find the sweet spot to calibrate my CV with LB based on the number of shifts
", "votes": 0, "replies": [] }, { "id": 460285, "author_name": "Scirpus", "author_url": "", "post_date": "2019-01-23T11:11:54.713000", "content": "I used welch using a point shift (np.arange(100,1000,100))and got 1.62 using Genetic Programming and 1.67 with lgb - might be useful for an ensemble
", "votes": 0, "replies": [] } ] }, { "id": 460784, "author_name": "Scirpus", "author_url": "", "post_date": "2019-01-24T12:17:15.017000", "content": "Andrew's Script + GP == 133 features for a score of 1.461
", "votes": 4, "replies": [ { "id": 460796, "author_name": "Abhishek Thakur", "author_url": "", "post_date": "2019-01-24T12:46:41.447000", "content": "gaussian process?
", "votes": 1, "replies": [] }, { "id": 460797, "author_name": "Scirpus", "author_url": "", "post_date": "2019-01-24T12:48:52.273000", "content": "Apologies, GP = Genetic Programming
", "votes": 1, "replies": [] }, { "id": 460800, "author_name": "Abhishek Thakur", "author_url": "", "post_date": "2019-01-24T12:52:17.163000", "content": "saw the previous post later. good stuff!
", "votes": 1, "replies": [] }, { "id": 460802, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2019-01-24T13:03:40.190000", "content": "@Abhishek, GP is so 2nd nature to @Scirpus he thinks we all hear Genetic Programming when we see GP instead of gaussian process or general practitioner :-)
", "votes": 2, "replies": [] }, { "id": 460805, "author_name": "Scirpus", "author_url": "", "post_date": "2019-01-24T13:07:05.463000", "content": "LOL so true ;)
", "votes": 2, "replies": [] }, { "id": 509019, "author_name": "Eyas Taifour", "author_url": "", "post_date": "2019-04-07T07:56:59.527000", "content": "what is Andy's script - sorry for the newb questoin
", "votes": 0, "replies": [] }, { "id": 509292, "author_name": "Ben Nye", "author_url": "", "post_date": "2019-04-07T16:26:32.577000", "content": "I expect that they mean the rather popular kernel by Andrew Lukyanenko , see https://www.kaggle.com/artgor/earthquakes-fe-more-features-and-samples
", "votes": 0, "replies": [] } ] }, { "id": 455698, "author_name": "", "author_url": "", "post_date": "2019-01-14T11:42:24.620000", "content": "", "votes": 0, "replies": [] }, { "id": 464988, "author_name": "QiPANDA", "author_url": "", "post_date": "2019-02-02T00:52:50.840000", "content": "Thanks
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "455565": "Since this is a feature building competition, it would be a good idea to know how many features is everyone using :)\n\n\nApprox. 60 features / LB 1.505\n\n", "456160": "currently, I am using just few aggregated features, as following:\n- mean, std, var, min, max, abs_max / segment;\n- A0-A10: being the FFT module values for CC and first 10 harmonics; R1-R5: being the real components of 1st 5 harmonics;\nTotally: 21 features / LB: 1.884 (very poor).\n\nI am also calculated 10 times more segments than the 4194 resulting by simply dividing the train data in 150k segments (I simply shift segments with a shift = 1/10 * segment, a technique for data multiplication we used to use in the past: we call it 'shifting aperture').\n\nIt looks that my validation error is quite close to the LB score, but the train error is much smaller so I do have a major problem with my model, a lot to improve to it.", "460784": "Andrew's Script + GP == 133 features for a score of 1.461", "455698": "", "464988": "Thanks" } }