So here's the thing-
\n\nOn one hand, it seems like a good number of participants are using very complex LSTMs hoping to \"get lucky\" with training
\n\nOn the other hand, I'm at LB 0.685 with an extremely simple 6 feature LGBM + pre/post processing (as per T. Vantuch's paper, plus some tricks)
\n\nThis looks surprising to me, almost hitting LB scores of top public kernels using a ridiculously simpler approach... which reinforces my hypothesis that some kind of LSTM lottery is being played >.<
\n\nWhat are your thoughts? Is it just random seed magic?
\n\nPS: Does anyone with an original approach fancy teaming up? (above public kernel score if using deep learning obviously)
\n\n(edit: 0.688 now, trying to beat the infamous 0.694 :O)
", "messages": [ { "id": "484394", "postDate": "03/05/2019 23:40:58", "content": "So here's the thing-
\n\nOn one hand, it seems like a good number of participants are using very complex LSTMs hoping to \"get lucky\" with training
\n\nOn the other hand, I'm at LB 0.685 with an extremely simple 6 feature LGBM + pre/post processing (as per T. Vantuch's paper, plus some tricks)
\n\nThis looks surprising to me, almost hitting LB scores of top public kernels using a ridiculously simpler approach... which reinforces my hypothesis that some kind of LSTM lottery is being played >.<
\n\nWhat are your thoughts? Is it just random seed magic?
\n\nPS: Does anyone with an original approach fancy teaming up? (above public kernel score if using deep learning obviously)
\n\n(edit: 0.688 now, trying to beat the infamous 0.694 :O)
", "rawMarkdown": "So here's the thing- \n\nOn one hand, it seems like a good number of participants are using very complex LSTMs hoping to \"get lucky\" with training\n\nOn the other hand, I'm at LB 0.685 with an extremely simple 6 feature LGBM + pre/post processing (as per T. Vantuch's paper, plus some tricks)\n\nThis looks surprising to me, almost hitting LB scores of top public kernels using a ridiculously simpler approach... which reinforces my hypothesis that some kind of LSTM lottery is being played >.<\n\nWhat are your thoughts? Is it just random seed magic?\n\nPS: Does anyone with an original approach fancy teaming up? (above public kernel score if using deep learning obviously)\n\n(edit: 0.688 now, trying to beat the infamous 0.694 :O)", "votes": null }, { "id": "484551", "postDate": "03/06/2019 07:04:51", "content": "How about your local cv?
\n\n6 feature? some statistics values of peaks?
\n\nDo you think wavelet denoising is useful?
", "rawMarkdown": "How about your local cv?\n\n6 feature? some statistics values of peaks?\n\nDo you think wavelet denoising is useful?", "votes": null }, { "id": "484732", "postDate": "03/06/2019 12:00:16", "content": "\n\n\nHow about your local cv?
\n
0.694 CV for 0.685 LB, stable
\n\n\n\n\n6 feature? some statistics values of peaks?
\n
Yes 6 peak statistics
\n\n\n\n\nDo you think wavelet denoising is useful?
\n
I think so, at least for me it helped alot
\n\nAre you also using tree methods?
", "rawMarkdown": "> How about your local cv?\n\n 0.694 CV for 0.685 LB, stable\n\n\n> 6 feature? some statistics values of peaks?\n\n Yes 6 peak statistics\n\n\n> Do you think wavelet denoising is useful?\n\n I think so, at least for me it helped alot\n\n\nAre you also using tree methods?", "votes": null }, { "id": "484786", "postDate": "03/06/2019 13:36:48", "content": "Hi Ganfear, I agree that top public-kernel neural networks with 10,000 or even 100,000 parameters may be a little overkill for this problem (anyway the kernel is great). I tried a smaller network (around 3000 parameters), and it can still fit training data without any problems ... (actually still overfit though) ...
\n\nPerhaps complex network design is still appropriate if we want diversity and combine them in several versions (ensemble) to reduce the variance.
", "rawMarkdown": "Hi Ganfear, I agree that top public-kernel neural networks with 10,000 or even 100,000 parameters may be a little overkill for this problem (anyway the kernel is great). I tried a smaller network (around 3000 parameters), and it can still fit training data without any problems ... (actually still overfit though) ... \n\nPerhaps complex network design is still appropriate if we want diversity and combine them in several versions (ensemble) to reduce the variance.", "votes": null }, { "id": "484931", "postDate": "03/06/2019 16:51:54", "content": "Hi, What is the meaning of preprocessing?
\n\nI understand the preprocessing is the human's judgement based on the experience or knowledge of PD.\nSo, after reading the paper and catch the important patterns of PD, someone can fix the predicted label.
\n\nIs this real meaning of your pre-processing?
", "rawMarkdown": "Hi, What is the meaning of preprocessing?\n\nI understand the preprocessing is the human's judgement based on the experience or knowledge of PD.\nSo, after reading the paper and catch the important patterns of PD, someone can fix the predicted label. \n\nIs this real meaning of your pre-processing?", "votes": null }, { "id": "484964", "postDate": "03/06/2019 17:52:01", "content": "Hi Neuron Engineer,
\n\n\n\n\nPerhaps complex network design is still appropriate if we want diversity and combine them in several versions (ensemble) to reduce the variance.
\n
Makes sense, yet it would rely heavily on... luck? Thanks for sharing your thoughts!
", "rawMarkdown": "Hi Neuron Engineer,\n\n> Perhaps complex network design is still appropriate if we want diversity and combine them in several versions (ensemble) to reduce the variance.\n\nMakes sense, yet it would rely heavily on... luck? Thanks for sharing your thoughts!", "votes": null }, { "id": "484969", "postDate": "03/06/2019 17:58:14", "content": "Hi YouHan,
\n\nNo, my preprocessing is a combination of:\n- False Peak Suppression as mentioned in the paper + minor tricks\n- Wavelet denoising
\n\nI think you'd really need to be an expert to label test samples by hand! Maybe someone at the top tried it?
", "rawMarkdown": "Hi YouHan,\n\nNo, my preprocessing is a combination of:\n- False Peak Suppression as mentioned in the paper + minor tricks\n- Wavelet denoising\n\nI think you'd really need to be an expert to label test samples by hand! Maybe someone at the top tried it?", "votes": null }, { "id": "485047", "postDate": "03/06/2019 21:14:41", "content": "\n\n\nOn the other hand, I'm at LB 0.685 with an extremely simple 6 feature LGBM + pre/post processing (as per T. Vantuch's paper, plus some tricks)
\n
Excuse me, what paper? Is there any kernel about that?
", "rawMarkdown": "> On the other hand, I'm at LB 0.685 with an extremely simple 6 feature LGBM + pre/post processing (as per T. Vantuch's paper, plus some tricks)\n\nExcuse me, what paper? Is there any kernel about that?", "votes": null }, { "id": "485060", "postDate": "03/06/2019 21:24:17", "content": "I believe, this one \nhttps://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/75771#447236
", "rawMarkdown": "I believe, this one \nhttps://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/75771#447236", "votes": null }, { "id": "485179", "postDate": "03/07/2019 03:14:05", "content": "How do you remove sysmetric peaks?
", "rawMarkdown": "How do you remove sysmetric peaks?", "votes": null }, { "id": "485368", "postDate": "03/07/2019 09:52:41", "content": "Not exactly luck, more law of large numbers?
", "rawMarkdown": "Not exactly luck, more law of large numbers?", "votes": null }, { "id": "485659", "postDate": "03/07/2019 18:35:43", "content": "As HarshitMehta mentioned below - https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/75771#447236
", "rawMarkdown": "As HarshitMehta mentioned below - https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/75771#447236", "votes": null }, { "id": "486398", "postDate": "03/08/2019 18:05:02", "content": "Is your offer about teaming up still actual?
", "rawMarkdown": "Is your offer about teaming up still actual?", "votes": null }, { "id": "486520", "postDate": "03/08/2019 23:04:52", "content": "Yes- just sent you PM
", "rawMarkdown": "Yes- just sent you PM", "votes": null }, { "id": "486696", "postDate": "03/09/2019 08:44:13", "content": "yea, this is a possible pitfall of these competitions...i hope this one is not that case so much, bud still crossing fingers for approaches like yours. 0.688 is great score! would you be willing to share it after the competition trough the kernel?
", "rawMarkdown": "yea, this is a possible pitfall of these competitions...i hope this one is not that case so much, bud still crossing fingers for approaches like yours. 0.688 is great score! would you be willing to share it after the competition trough the kernel?", "votes": null }, { "id": "486888", "postDate": "03/09/2019 15:55:31", "content": "Hi Tomas! I also have a request, would it be possible to share (or hint) how you label the PD cases after the competition? By inspecting many and many waveforms without seeing any obvious patterns, I am also super curious to understand the labeling method.
", "rawMarkdown": "Hi Tomas! I also have a request, would it be possible to share (or hint) how you label the PD cases after the competition? By inspecting many and many waveforms without seeing any obvious patterns, I am also super curious to understand the labeling method.", "votes": null }, { "id": "486910", "postDate": "03/09/2019 16:47:39", "content": "That's amazing result GanFear. The best i can get using decision tree is .660 lb and .716 cv after taking help from couple of kernels for extracting features and blending them with my features. i am using 96 features.
", "rawMarkdown": "That's amazing result GanFear. The best i can get using decision tree is .660 lb and .716 cv after taking help from couple of kernels for extracting features and blending them with my features. i am using 96 features.", "votes": null }, { "id": "486928", "postDate": "03/09/2019 17:24:10", "content": "> it seems like a good number of participants are using very complex LSTMs hoping to \"get lucky\" with training
\n\nYeah, let's cross fingers and play ML roulette! :D
", "rawMarkdown": "> it seems like a good number of participants are using very complex LSTMs hoping to \"get lucky\" with training\n\nYeah, let's cross fingers and play ML roulette! :D", "votes": null }, { "id": "486996", "postDate": "03/09/2019 21:37:27", "content": "Hi Tomas,\nSure I'd be happy to share my approach once the competition is over :)
", "rawMarkdown": "Hi Tomas,\nSure I'd be happy to share my approach once the competition is over :)", "votes": null }, { "id": "486999", "postDate": "03/09/2019 21:42:39", "content": "Thanks! I'd bet you can improve those 0.660 by removing a couple of less important features- at least it worked for me (started with 50+ features I'm down to 6 now)
", "rawMarkdown": "Thanks! I'd bet you can improve those 0.660 by removing a couple of less important features- at least it worked for me (started with 50+ features I'm down to 6 now)", "votes": null }, { "id": "487001", "postDate": "03/09/2019 21:45:36", "content": "And hope the ML roulette doesn't cause a huge LB shakeup! :O
", "rawMarkdown": "And hope the ML roulette doesn't cause a huge LB shakeup! :O", "votes": null }, { "id": "488895", "postDate": "03/13/2019 06:43:28", "content": "Yeah, I hope too! :D Probably I (and maybe others) overfitted to LB. This can lead to sad consequences.\nRecently I've read about the super failure in another competition:
\n\nhttps://www.kaggle.com/c/santander-value-prediction-challenge/discussion/63753
", "rawMarkdown": "Yeah, I hope too! :D Probably I (and maybe others) overfitted to LB. This can lead to sad consequences.\nRecently I've read about the super failure in another competition: \n\nhttps://www.kaggle.com/c/santander-value-prediction-challenge/discussion/63753", "votes": null } ], "comments": [ { "id": 484551, "author_name": "blackboards", "author_url": "", "post_date": "03/06/2019 07:04:51", "content": "How about your local cv?
\n\n6 feature? some statistics values of peaks?
\n\nDo you think wavelet denoising is useful?
", "votes": null, "replies": [ { "id": 484732, "author_name": "ganfear", "author_url": "", "post_date": "03/06/2019 12:00:16", "content": "\n\n\nHow about your local cv?
\n
0.694 CV for 0.685 LB, stable
\n\n\n\n\n6 feature? some statistics values of peaks?
\n
Yes 6 peak statistics
\n\n\n\n\nDo you think wavelet denoising is useful?
\n
I think so, at least for me it helped alot
\n\nAre you also using tree methods?
", "votes": null, "replies": [] } ] }, { "id": 484786, "author_name": "ratthachat", "author_url": "", "post_date": "03/06/2019 13:36:48", "content": "Hi Ganfear, I agree that top public-kernel neural networks with 10,000 or even 100,000 parameters may be a little overkill for this problem (anyway the kernel is great). I tried a smaller network (around 3000 parameters), and it can still fit training data without any problems ... (actually still overfit though) ...
\n\nPerhaps complex network design is still appropriate if we want diversity and combine them in several versions (ensemble) to reduce the variance.
", "votes": null, "replies": [ { "id": 484964, "author_name": "ganfear", "author_url": "", "post_date": "03/06/2019 17:52:01", "content": "Hi Neuron Engineer,
\n\n\n\n\nPerhaps complex network design is still appropriate if we want diversity and combine them in several versions (ensemble) to reduce the variance.
\n
Makes sense, yet it would rely heavily on... luck? Thanks for sharing your thoughts!
", "votes": null, "replies": [] }, { "id": 485368, "author_name": "proef2", "author_url": "", "post_date": "03/07/2019 09:52:41", "content": "Not exactly luck, more law of large numbers?
", "votes": null, "replies": [] } ] }, { "id": 484931, "author_name": "youhanlee", "author_url": "", "post_date": "03/06/2019 16:51:54", "content": "Hi, What is the meaning of preprocessing?
\n\nI understand the preprocessing is the human's judgement based on the experience or knowledge of PD.\nSo, after reading the paper and catch the important patterns of PD, someone can fix the predicted label.
\n\nIs this real meaning of your pre-processing?
", "votes": null, "replies": [ { "id": 484969, "author_name": "ganfear", "author_url": "", "post_date": "03/06/2019 17:58:14", "content": "Hi YouHan,
\n\nNo, my preprocessing is a combination of:\n- False Peak Suppression as mentioned in the paper + minor tricks\n- Wavelet denoising
\n\nI think you'd really need to be an expert to label test samples by hand! Maybe someone at the top tried it?
", "votes": null, "replies": [] } ] }, { "id": 485047, "author_name": "davids1992", "author_url": "", "post_date": "03/06/2019 21:14:41", "content": "\n\n\nOn the other hand, I'm at LB 0.685 with an extremely simple 6 feature LGBM + pre/post processing (as per T. Vantuch's paper, plus some tricks)
\n
Excuse me, what paper? Is there any kernel about that?
", "votes": null, "replies": [ { "id": 485060, "author_name": "harshit92", "author_url": "", "post_date": "03/06/2019 21:24:17", "content": "I believe, this one \nhttps://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/75771#447236
", "votes": null, "replies": [] } ] }, { "id": 485179, "author_name": "gmhost", "author_url": "", "post_date": "03/07/2019 03:14:05", "content": "How do you remove sysmetric peaks?
", "votes": null, "replies": [ { "id": 485659, "author_name": "ganfear", "author_url": "", "post_date": "03/07/2019 18:35:43", "content": "As HarshitMehta mentioned below - https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/75771#447236
", "votes": null, "replies": [] } ] }, { "id": 486398, "author_name": "raf123", "author_url": "", "post_date": "03/08/2019 18:05:02", "content": "Is your offer about teaming up still actual?
", "votes": null, "replies": [ { "id": 486520, "author_name": "ganfear", "author_url": "", "post_date": "03/08/2019 23:04:52", "content": "Yes- just sent you PM
", "votes": null, "replies": [] } ] }, { "id": 486696, "author_name": "tvantuch", "author_url": "", "post_date": "03/09/2019 08:44:13", "content": "yea, this is a possible pitfall of these competitions...i hope this one is not that case so much, bud still crossing fingers for approaches like yours. 0.688 is great score! would you be willing to share it after the competition trough the kernel?
", "votes": null, "replies": [ { "id": 486888, "author_name": "ratthachat", "author_url": "", "post_date": "03/09/2019 15:55:31", "content": "Hi Tomas! I also have a request, would it be possible to share (or hint) how you label the PD cases after the competition? By inspecting many and many waveforms without seeing any obvious patterns, I am also super curious to understand the labeling method.
", "votes": null, "replies": [] }, { "id": 486996, "author_name": "ganfear", "author_url": "", "post_date": "03/09/2019 21:37:27", "content": "Hi Tomas,\nSure I'd be happy to share my approach once the competition is over :)
", "votes": null, "replies": [] } ] }, { "id": 486910, "author_name": "harshit92", "author_url": "", "post_date": "03/09/2019 16:47:39", "content": "That's amazing result GanFear. The best i can get using decision tree is .660 lb and .716 cv after taking help from couple of kernels for extracting features and blending them with my features. i am using 96 features.
", "votes": null, "replies": [ { "id": 486999, "author_name": "ganfear", "author_url": "", "post_date": "03/09/2019 21:42:39", "content": "Thanks! I'd bet you can improve those 0.660 by removing a couple of less important features- at least it worked for me (started with 50+ features I'm down to 6 now)
", "votes": null, "replies": [] } ] }, { "id": 486928, "author_name": "sergeyzlobin", "author_url": "", "post_date": "03/09/2019 17:24:10", "content": "> it seems like a good number of participants are using very complex LSTMs hoping to \"get lucky\" with training
\n\nYeah, let's cross fingers and play ML roulette! :D
", "votes": null, "replies": [ { "id": 487001, "author_name": "ganfear", "author_url": "", "post_date": "03/09/2019 21:45:36", "content": "And hope the ML roulette doesn't cause a huge LB shakeup! :O
", "votes": null, "replies": [] }, { "id": 488895, "author_name": "sergeyzlobin", "author_url": "", "post_date": "03/13/2019 06:43:28", "content": "Yeah, I hope too! :D Probably I (and maybe others) overfitted to LB. This can lead to sad consequences.\nRecently I've read about the super failure in another competition:
\n\nhttps://www.kaggle.com/c/santander-value-prediction-challenge/discussion/63753
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "484394": "So here's the thing- \n\nOn one hand, it seems like a good number of participants are using very complex LSTMs hoping to \"get lucky\" with training\n\nOn the other hand, I'm at LB 0.685 with an extremely simple 6 feature LGBM + pre/post processing (as per T. Vantuch's paper, plus some tricks)\n\nThis looks surprising to me, almost hitting LB scores of top public kernels using a ridiculously simpler approach... which reinforces my hypothesis that some kind of LSTM lottery is being played >.<\n\nWhat are your thoughts? Is it just random seed magic?\n\nPS: Does anyone with an original approach fancy teaming up? (above public kernel score if using deep learning obviously)\n\n(edit: 0.688 now, trying to beat the infamous 0.694 :O)", "484551": "How about your local cv?\n\n6 feature? some statistics values of peaks?\n\nDo you think wavelet denoising is useful?", "484732": "> How about your local cv?\n\n 0.694 CV for 0.685 LB, stable\n\n\n> 6 feature? some statistics values of peaks?\n\n Yes 6 peak statistics\n\n\n> Do you think wavelet denoising is useful?\n\n I think so, at least for me it helped alot\n\n\nAre you also using tree methods?", "484786": "Hi Ganfear, I agree that top public-kernel neural networks with 10,000 or even 100,000 parameters may be a little overkill for this problem (anyway the kernel is great). I tried a smaller network (around 3000 parameters), and it can still fit training data without any problems ... (actually still overfit though) ... \n\nPerhaps complex network design is still appropriate if we want diversity and combine them in several versions (ensemble) to reduce the variance.", "484931": "Hi, What is the meaning of preprocessing?\n\nI understand the preprocessing is the human's judgement based on the experience or knowledge of PD.\nSo, after reading the paper and catch the important patterns of PD, someone can fix the predicted label. \n\nIs this real meaning of your pre-processing?", "484964": "Hi Neuron Engineer,\n\n> Perhaps complex network design is still appropriate if we want diversity and combine them in several versions (ensemble) to reduce the variance.\n\nMakes sense, yet it would rely heavily on... luck? Thanks for sharing your thoughts!", "484969": "Hi YouHan,\n\nNo, my preprocessing is a combination of:\n- False Peak Suppression as mentioned in the paper + minor tricks\n- Wavelet denoising\n\nI think you'd really need to be an expert to label test samples by hand! Maybe someone at the top tried it?", "485047": "> On the other hand, I'm at LB 0.685 with an extremely simple 6 feature LGBM + pre/post processing (as per T. Vantuch's paper, plus some tricks)\n\nExcuse me, what paper? Is there any kernel about that?", "485060": "I believe, this one \nhttps://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/75771#447236", "485179": "How do you remove sysmetric peaks?", "485368": "Not exactly luck, more law of large numbers?", "485659": "As HarshitMehta mentioned below - https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/75771#447236", "486398": "Is your offer about teaming up still actual?", "486520": "Yes- just sent you PM", "486696": "yea, this is a possible pitfall of these competitions...i hope this one is not that case so much, bud still crossing fingers for approaches like yours. 0.688 is great score! would you be willing to share it after the competition trough the kernel?", "486888": "Hi Tomas! I also have a request, would it be possible to share (or hint) how you label the PD cases after the competition? By inspecting many and many waveforms without seeing any obvious patterns, I am also super curious to understand the labeling method.", "486910": "That's amazing result GanFear. The best i can get using decision tree is .660 lb and .716 cv after taking help from couple of kernels for extracting features and blending them with my features. i am using 96 features.", "486928": "> it seems like a good number of participants are using very complex LSTMs hoping to \"get lucky\" with training\n\nYeah, let's cross fingers and play ML roulette! :D", "486996": "Hi Tomas,\nSure I'd be happy to share my approach once the competition is over :)", "486999": "Thanks! I'd bet you can improve those 0.660 by removing a couple of less important features- at least it worked for me (started with 50+ features I'm down to 6 now)", "487001": "And hope the ML roulette doesn't cause a huge LB shakeup! :O", "488895": "Yeah, I hope too! :D Probably I (and maybe others) overfitted to LB. This can lead to sad consequences.\nRecently I've read about the super failure in another competition: \n\nhttps://www.kaggle.com/c/santander-value-prediction-challenge/discussion/63753" }, "source": "meta" }