Hello,
\n\nSorry if this a stupid question (I'm new to this) but I've been extracting the data from train.csv and plotted it, only to observe that the (supposed) seismic event happens way before the time_to_failure reaches zero. You can see that in the attached picture: the peak seismic activity is around 4500000 but the time reaches 0 around 5500000. Did I overlook something? Haven't I understood the requirements?
\n\nThanks in advance! :)
", "messages": [ { "id": "469127", "postDate": "02/10/2019 14:53:41", "content": "Hello,
\n\nSorry if this a stupid question (I'm new to this) but I've been extracting the data from train.csv and plotted it, only to observe that the (supposed) seismic event happens way before the time_to_failure reaches zero. You can see that in the attached picture: the peak seismic activity is around 4500000 but the time reaches 0 around 5500000. Did I overlook something? Haven't I understood the requirements?
\n\nThanks in advance! :)
", "rawMarkdown": "Hello,\n\nSorry if this a stupid question (I'm new to this) but I've been extracting the data from train.csv and plotted it, only to observe that the (supposed) seismic event happens way before the time_to_failure reaches zero. You can see that in the attached picture: the peak seismic activity is around 4500000 but the time reaches 0 around 5500000. Did I overlook something? Haven't I understood the requirements?\n\nThanks in advance! :)", "votes": null }, { "id": "469327", "postDate": "02/11/2019 01:55:05", "content": "Essentially the large seismic event is not the earthquake. It is when time_to_failure = 0 that the lab earthquake occurs. The two threads from Bertrand can help with understanding the data a bit more (I'm having trouble with the hyperlink in the comment box, but they are the top posts in discussion).
", "rawMarkdown": "Essentially the large seismic event is not the earthquake. It is when time_to_failure = 0 that the lab earthquake occurs. The two threads from Bertrand can help with understanding the data a bit more (I'm having trouble with the hyperlink in the comment box, but they are the top posts in discussion).", "votes": null }, { "id": "469468", "postDate": "02/11/2019 09:20:51", "content": "So basically, the earthquake occurs in the lab, but it is not registered by the signal? It just doesn't show up on the seismograph?
", "rawMarkdown": "So basically, the earthquake occurs in the lab, but it is not registered by the signal? It just doesn't show up on the seismograph?", "votes": null }, { "id": "469507", "postDate": "02/11/2019 10:54:58", "content": "Looking at all the training earthquakes, there is a constant deviation of 0.3s, between the high signal and the zero time to failure. During that 0.3 seconds, the signal is very low and similar to the following 1 second, as if the earthquake had really occured.\nI tried to \"change\" the timetofailure in my training data (removing first 0.3 seconds of timetofailure data).\nThen the training CV score is much better (nearly 0.1 second better). But I don't manage to improve the score on the real validation data (nor on the public LB).
", "rawMarkdown": "Looking at all the training earthquakes, there is a constant deviation of 0.3s, between the high signal and the zero time to failure. During that 0.3 seconds, the signal is very low and similar to the following 1 second, as if the earthquake had really occured.\nI tried to \"change\" the timetofailure in my training data (removing first 0.3 seconds of timetofailure data).\nThen the training CV score is much better (nearly 0.1 second better). But I don't manage to improve the score on the real validation data (nor on the public LB).", "votes": null }, { "id": "469539", "postDate": "02/11/2019 11:49:22", "content": "I'm throwing out last 0.3 seconds before each earthquake and it improves my score.
", "rawMarkdown": "I'm throwing out last 0.3 seconds before each earthquake and it improves my score.", "votes": null }, { "id": "469621", "postDate": "02/11/2019 14:57:42", "content": "From the paper 'Machine Learning Predicts Laboratory Earthquakes', in Figure 3 the authors mention \"... a classic, impulsive acoustic emission observed in the critically stressed region just preceding failure\", which I have been taking as an explanation of the time delta that you have brought up here.
", "rawMarkdown": "From the paper 'Machine Learning Predicts Laboratory Earthquakes', in Figure 3 the authors mention \"... a classic, impulsive acoustic emission observed in the critically stressed region just preceding failure\", which I have been taking as an explanation of the time delta that you have brought up here.", "votes": null }, { "id": "469787", "postDate": "02/11/2019 20:00:56", "content": "Im gonna try that and update after
", "rawMarkdown": "Im gonna try that and update after", "votes": null }, { "id": "471226", "postDate": "02/14/2019 06:54:10", "content": "I believe that they use acoustic emission sensor, not a seismograph.
", "rawMarkdown": "I believe that they use acoustic emission sensor, not a seismograph.", "votes": null }, { "id": "473580", "postDate": "02/18/2019 08:09:26", "content": "What really puzzles me is that there seems to be a constant time ( 0.3 sec ) between the highest signal and the zero time to failure in the training data. If this is correct, it would mean that the true lab earthquake in all experiments performed happens exactly 0.3 sec after the pre-earthquake high signal occured. I find this very unlikely. It would be helpful if maybe someone could explain that finding from a experimental physics point of view (and not from a data science perspective).
", "rawMarkdown": "What really puzzles me is that there seems to be a constant time ( 0.3 sec ) between the highest signal and the zero time to failure in the training data. If this is correct, it would mean that the true lab earthquake in all experiments performed happens exactly 0.3 sec after the pre-earthquake high signal occured. I find this very unlikely. It would be helpful if maybe someone could explain that finding from a experimental physics point of view (and not from a data science perspective).", "votes": null }, { "id": "527801", "postDate": "05/06/2019 10:26:48", "content": "They are not constant. For every of 16 \"earthquakes\" delay value equals accordingly to\n0.32029613\n0.31119797\n0.31779757\n0.31079901\n0.31859753\n0.32055664\n0.31616211\n0.31494141\n0.31689453\n0.31860352\n0.31884766
", "rawMarkdown": "They are not constant. For every of 16 \"earthquakes\" delay value equals accordingly to\n0.32029613\n0.31119797\n0.31779757\n0.31079901\n0.31859753\n0.32055664\n0.31616211\n0.31494141\n0.31689453\n0.31860352\n0.31884766", "votes": null }, { "id": "527806", "postDate": "05/06/2019 10:39:47", "content": "Not constant, but std is very low still at 0.003.
", "rawMarkdown": "Not constant, but std is very low still at 0.003.", "votes": null } ], "comments": [ { "id": 469327, "author_name": "", "author_url": "", "post_date": "02/11/2019 01:55:05", "content": "Essentially the large seismic event is not the earthquake. It is when time_to_failure = 0 that the lab earthquake occurs. The two threads from Bertrand can help with understanding the data a bit more (I'm having trouble with the hyperlink in the comment box, but they are the top posts in discussion).
", "votes": null, "replies": [ { "id": 469468, "author_name": "tuddy18", "author_url": "", "post_date": "02/11/2019 09:20:51", "content": "So basically, the earthquake occurs in the lab, but it is not registered by the signal? It just doesn't show up on the seismograph?
", "votes": null, "replies": [] }, { "id": 469621, "author_name": "", "author_url": "", "post_date": "02/11/2019 14:57:42", "content": "From the paper 'Machine Learning Predicts Laboratory Earthquakes', in Figure 3 the authors mention \"... a classic, impulsive acoustic emission observed in the critically stressed region just preceding failure\", which I have been taking as an explanation of the time delta that you have brought up here.
", "votes": null, "replies": [] }, { "id": 471226, "author_name": "petya5q", "author_url": "", "post_date": "02/14/2019 06:54:10", "content": "I believe that they use acoustic emission sensor, not a seismograph.
", "votes": null, "replies": [] } ] }, { "id": 469507, "author_name": "zidmie", "author_url": "", "post_date": "02/11/2019 10:54:58", "content": "Looking at all the training earthquakes, there is a constant deviation of 0.3s, between the high signal and the zero time to failure. During that 0.3 seconds, the signal is very low and similar to the following 1 second, as if the earthquake had really occured.\nI tried to \"change\" the timetofailure in my training data (removing first 0.3 seconds of timetofailure data).\nThen the training CV score is much better (nearly 0.1 second better). But I don't manage to improve the score on the real validation data (nor on the public LB).
", "votes": null, "replies": [ { "id": 469539, "author_name": "elvenmonk", "author_url": "", "post_date": "02/11/2019 11:49:22", "content": "I'm throwing out last 0.3 seconds before each earthquake and it improves my score.
", "votes": null, "replies": [] }, { "id": 469787, "author_name": "tuddy18", "author_url": "", "post_date": "02/11/2019 20:00:56", "content": "Im gonna try that and update after
", "votes": null, "replies": [] } ] }, { "id": 473580, "author_name": "udoschmitt", "author_url": "", "post_date": "02/18/2019 08:09:26", "content": "What really puzzles me is that there seems to be a constant time ( 0.3 sec ) between the highest signal and the zero time to failure in the training data. If this is correct, it would mean that the true lab earthquake in all experiments performed happens exactly 0.3 sec after the pre-earthquake high signal occured. I find this very unlikely. It would be helpful if maybe someone could explain that finding from a experimental physics point of view (and not from a data science perspective).
", "votes": null, "replies": [ { "id": 527801, "author_name": "vbelobragin", "author_url": "", "post_date": "05/06/2019 10:26:48", "content": "They are not constant. For every of 16 \"earthquakes\" delay value equals accordingly to\n0.32029613\n0.31119797\n0.31779757\n0.31079901\n0.31859753\n0.32055664\n0.31616211\n0.31494141\n0.31689453\n0.31860352\n0.31884766
", "votes": null, "replies": [] }, { "id": 527806, "author_name": "cpmpml", "author_url": "", "post_date": "05/06/2019 10:39:47", "content": "Not constant, but std is very low still at 0.003.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "469127": "Hello,\n\nSorry if this a stupid question (I'm new to this) but I've been extracting the data from train.csv and plotted it, only to observe that the (supposed) seismic event happens way before the time_to_failure reaches zero. You can see that in the attached picture: the peak seismic activity is around 4500000 but the time reaches 0 around 5500000. Did I overlook something? Haven't I understood the requirements?\n\nThanks in advance! :)", "469327": "Essentially the large seismic event is not the earthquake. It is when time_to_failure = 0 that the lab earthquake occurs. The two threads from Bertrand can help with understanding the data a bit more (I'm having trouble with the hyperlink in the comment box, but they are the top posts in discussion).", "469468": "So basically, the earthquake occurs in the lab, but it is not registered by the signal? It just doesn't show up on the seismograph?", "469507": "Looking at all the training earthquakes, there is a constant deviation of 0.3s, between the high signal and the zero time to failure. During that 0.3 seconds, the signal is very low and similar to the following 1 second, as if the earthquake had really occured.\nI tried to \"change\" the timetofailure in my training data (removing first 0.3 seconds of timetofailure data).\nThen the training CV score is much better (nearly 0.1 second better). But I don't manage to improve the score on the real validation data (nor on the public LB).", "469539": "I'm throwing out last 0.3 seconds before each earthquake and it improves my score.", "469621": "From the paper 'Machine Learning Predicts Laboratory Earthquakes', in Figure 3 the authors mention \"... a classic, impulsive acoustic emission observed in the critically stressed region just preceding failure\", which I have been taking as an explanation of the time delta that you have brought up here.", "469787": "Im gonna try that and update after", "471226": "I believe that they use acoustic emission sensor, not a seismograph.", "473580": "What really puzzles me is that there seems to be a constant time ( 0.3 sec ) between the highest signal and the zero time to failure in the training data. If this is correct, it would mean that the true lab earthquake in all experiments performed happens exactly 0.3 sec after the pre-earthquake high signal occured. I find this very unlikely. It would be helpful if maybe someone could explain that finding from a experimental physics point of view (and not from a data science perspective).", "527801": "They are not constant. For every of 16 \"earthquakes\" delay value equals accordingly to\n0.32029613\n0.31119797\n0.31779757\n0.31079901\n0.31859753\n0.32055664\n0.31616211\n0.31494141\n0.31689453\n0.31860352\n0.31884766", "527806": "Not constant, but std is very low still at 0.003." }, "source": "meta" }