Regarding VSB partial fault detection, it is found by many top-score participants that the denoised version can efficiently increase predictive performance for neural networks (Note : not tree models) in private test set. So our team (credit : @putalay) think that by making this dataset available should benefit and save the time for our community.
\n\nWe apply train/test parquet files with the method described by Jack @jackvial : https://www.kaggle.com/jackvial/dwt-signal-denoising
\n\nif you are interested, please import this data into your kernel : \nhttps://www.kaggle.com/thaikeras/vsb-wavelet-denoised/
\n\nYou can use this data directly in place of the original train.parq / test.parq
\n\n1) by applying this data version to our top public kernels (e.g. @braquino Bruno's and @tarunpaparaju Tarun's), we immediately got the privateLB results around 0.640 - 0.670 ... So, in the first place, people who believe in this denoising method already won medals.
\n\n2) However, it is certainly not easy to decide and believe in this method since direct application will reduce public LB score to 0.630-0.670
\n\n3) In order to also get a good public score (as well as private scores) we need to do more, for examples by some insights of our participants :
\n\nJustify the validation of the method (using adversarial validation) as mentioned by @yukinkgwa here
Apply ensembles : as used by many top scorers
Apply justified complicated pre-processing techniques like @vhessel here
Use a good probed data in addition to a training data as @bigswimatom in https://www.kaggle.com/bigswimatom/5-fold-lstm-attention-stateful-metrics-with-exp\n(I empirically found that stateful metric doesn't help much as the probed data)
or if you are able to find a killer feature like @tw1994 Tang's here
4) This denoised data seems cannot improve the performance of the tree based method. (I try to use it in https://www.kaggle.com/qinhui1999/handmade-features-0685-private-score )\nBut @qinhui1999 Hui’s pseudo labelling method could also improve the publicLB in the DL kernel.
\n\nFinal note to @sheriytm my friend, I tried re-implemented many methods shared by our top participants, but the single most important and effective factor (exclude more complicated techniques mentioned above) seems to be the denoised dataset here, so I share it here as promised to you :) .
", "messages": [ { "id": "502815", "postDate": "03/29/2019 05:19:17", "content": "Regarding VSB partial fault detection, it is found by many top-score participants that the denoised version can efficiently increase predictive performance for neural networks (Note : not tree models) in private test set. So our team (credit : @putalay) think that by making this dataset available should benefit and save the time for our community.
\n\nWe apply train/test parquet files with the method described by Jack @jackvial : https://www.kaggle.com/jackvial/dwt-signal-denoising
\n\nif you are interested, please import this data into your kernel : \nhttps://www.kaggle.com/thaikeras/vsb-wavelet-denoised/
\n\nYou can use this data directly in place of the original train.parq / test.parq
\n\n1) by applying this data version to our top public kernels (e.g. @braquino Bruno's and @tarunpaparaju Tarun's), we immediately got the privateLB results around 0.640 - 0.670 ... So, in the first place, people who believe in this denoising method already won medals.
\n\n2) However, it is certainly not easy to decide and believe in this method since direct application will reduce public LB score to 0.630-0.670
\n\n3) In order to also get a good public score (as well as private scores) we need to do more, for examples by some insights of our participants :
\n\nJustify the validation of the method (using adversarial validation) as mentioned by @yukinkgwa here
Apply ensembles : as used by many top scorers
Apply justified complicated pre-processing techniques like @vhessel here
Use a good probed data in addition to a training data as @bigswimatom in https://www.kaggle.com/bigswimatom/5-fold-lstm-attention-stateful-metrics-with-exp\n(I empirically found that stateful metric doesn't help much as the probed data)
or if you are able to find a killer feature like @tw1994 Tang's here
4) This denoised data seems cannot improve the performance of the tree based method. (I try to use it in https://www.kaggle.com/qinhui1999/handmade-features-0685-private-score )\nBut @qinhui1999 Hui’s pseudo labelling method could also improve the publicLB in the DL kernel.
\n\nFinal note to @sheriytm my friend, I tried re-implemented many methods shared by our top participants, but the single most important and effective factor (exclude more complicated techniques mentioned above) seems to be the denoised dataset here, so I share it here as promised to you :) .
", "rawMarkdown": "# Introduction\n\nRegarding VSB partial fault detection, it is found by many top-score participants that the denoised version can efficiently increase predictive performance **for neural networks** (Note : not tree models) in private test set. So our team (credit : @putalay) think that by making this dataset available should benefit and save the time for our community.\n\nWe apply train/test parquet files with the method described by Jack @jackvial : https://www.kaggle.com/jackvial/dwt-signal-denoising\n\n## Dataset URL\nif you are interested, please import this data into your kernel : \nhttps://www.kaggle.com/thaikeras/vsb-wavelet-denoised/\n\nYou can use this data directly in place of the original train.parq / test.parq\n\n## Effects and Notes on Performance Improvement\n\n1) by applying this data version to our top public kernels (e.g. @braquino Bruno's and @tarunpaparaju Tarun's), we immediately got the privateLB results around 0.640 - 0.670 ... **So, in the first place, people who believe in this denoising method already won medals.**\n\n2) However, it is certainly not easy to decide and believe in this method since direct application will reduce public LB score to 0.630-0.670\n\n3) In order to also get a good public score (as well as private scores) we need to do more, for examples by some insights of our participants : \n\n- Justify the validation of the method (using adversarial validation) as mentioned by @yukinkgwa [here](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85258)\n\n- Apply ensembles : as used by many top scorers\n\n- Apply justified complicated pre-processing techniques like @vhessel [here](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85170) \n\n- Use a good probed data in addition to a training data as @bigswimatom in https://www.kaggle.com/bigswimatom/5-fold-lstm-attention-stateful-metrics-with-exp\n(I empirically found that stateful metric doesn't help much as the probed data)\n\n- or if you are able to find a killer feature like @tw1994 Tang's [here](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/86616)\n\n\n4) This denoised data seems cannot improve the performance of the tree based method. (I try to use it in https://www.kaggle.com/qinhui1999/handmade-features-0685-private-score )\nBut @qinhui1999 Hui’s pseudo labelling method could also improve the publicLB in the DL kernel.\n\nFinal note to @sheriytm my friend, I tried re-implemented many methods shared by our top participants, but the single most important and effective factor (exclude more complicated techniques mentioned above) seems to be the denoised dataset here, so I share it here as promised to you :) .", "votes": null }, { "id": "502884", "postDate": "03/29/2019 07:38:41", "content": "Well done @ratthachat and thank you very much for generously sharing your findings. I find the problem addressed in this competion fascinating. My goal is to find time later on to build a simple model that can produce nearly as good a result as the top solutions. When and if I do that, I will be posting my updates here.
", "rawMarkdown": "Well done @ratthachat and thank you very much for generously sharing your findings. I find the problem addressed in this competion fascinating. My goal is to find time later on to build a simple model that can produce nearly as good a result as the top solutions. When and if I do that, I will be posting my updates here.", "votes": null }, { "id": "502932", "postDate": "03/29/2019 09:20:35", "content": "Thank you very much for citing my kernel.
", "rawMarkdown": "Thank you very much for citing my kernel.", "votes": null }, { "id": "503069", "postDate": "03/29/2019 12:39:30", "content": "Great work, @ratthachat ! Thanks for citing my kernel !
", "rawMarkdown": "Great work, @ratthachat ! Thanks for citing my kernel !", "votes": null }, { "id": "504122", "postDate": "03/31/2019 02:47:38", "content": "@ratthachat, I can confirm your observations that denoised data does not predictive performance of tree bsed models. I plugged you data denoised to my best model on the private LB i.e. my public Catboost kernel and it did not improve the private LB score of 0.62059 but scored the highest of all my models on th public LB with 0.71467. I will try it with my NN model when I resume experimentation on this data.
", "rawMarkdown": "ratthachat, I can confirm your observations that denoised data does not predictive performance of tree bsed models. I plugged you data denoised to [my best model](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85146) on the private LB i.e. [my public Catboost kernel](https://www.kaggle.com/sheriytm/5-fold-lstm-atten-fully-commented-original-cpu?scriptVersionId=10248597) and it did not improve the private LB score of 0.62059 but scored the highest of all my models on th public LB with 0.71467. I will try it with my NN model when I resume experimentation on this data.", "votes": null } ], "comments": [ { "id": 502884, "author_name": "sheriytm", "author_url": "", "post_date": "03/29/2019 07:38:41", "content": "Well done @ratthachat and thank you very much for generously sharing your findings. I find the problem addressed in this competion fascinating. My goal is to find time later on to build a simple model that can produce nearly as good a result as the top solutions. When and if I do that, I will be posting my updates here.
", "votes": null, "replies": [] }, { "id": 502932, "author_name": "bigswimatom", "author_url": "", "post_date": "03/29/2019 09:20:35", "content": "Thank you very much for citing my kernel.
", "votes": null, "replies": [] }, { "id": 503069, "author_name": "tarunpaparaju", "author_url": "", "post_date": "03/29/2019 12:39:30", "content": "Great work, @ratthachat ! Thanks for citing my kernel !
", "votes": null, "replies": [] }, { "id": 504122, "author_name": "sheriytm", "author_url": "", "post_date": "03/31/2019 02:47:38", "content": "@ratthachat, I can confirm your observations that denoised data does not predictive performance of tree bsed models. I plugged you data denoised to my best model on the private LB i.e. my public Catboost kernel and it did not improve the private LB score of 0.62059 but scored the highest of all my models on th public LB with 0.71467. I will try it with my NN model when I resume experimentation on this data.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "502815": "# Introduction\n\nRegarding VSB partial fault detection, it is found by many top-score participants that the denoised version can efficiently increase predictive performance **for neural networks** (Note : not tree models) in private test set. So our team (credit : @putalay) think that by making this dataset available should benefit and save the time for our community.\n\nWe apply train/test parquet files with the method described by Jack @jackvial : https://www.kaggle.com/jackvial/dwt-signal-denoising\n\n## Dataset URL\nif you are interested, please import this data into your kernel : \nhttps://www.kaggle.com/thaikeras/vsb-wavelet-denoised/\n\nYou can use this data directly in place of the original train.parq / test.parq\n\n## Effects and Notes on Performance Improvement\n\n1) by applying this data version to our top public kernels (e.g. @braquino Bruno's and @tarunpaparaju Tarun's), we immediately got the privateLB results around 0.640 - 0.670 ... **So, in the first place, people who believe in this denoising method already won medals.**\n\n2) However, it is certainly not easy to decide and believe in this method since direct application will reduce public LB score to 0.630-0.670\n\n3) In order to also get a good public score (as well as private scores) we need to do more, for examples by some insights of our participants : \n\n- Justify the validation of the method (using adversarial validation) as mentioned by @yukinkgwa [here](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85258)\n\n- Apply ensembles : as used by many top scorers\n\n- Apply justified complicated pre-processing techniques like @vhessel [here](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85170) \n\n- Use a good probed data in addition to a training data as @bigswimatom in https://www.kaggle.com/bigswimatom/5-fold-lstm-attention-stateful-metrics-with-exp\n(I empirically found that stateful metric doesn't help much as the probed data)\n\n- or if you are able to find a killer feature like @tw1994 Tang's [here](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/86616)\n\n\n4) This denoised data seems cannot improve the performance of the tree based method. (I try to use it in https://www.kaggle.com/qinhui1999/handmade-features-0685-private-score )\nBut @qinhui1999 Hui’s pseudo labelling method could also improve the publicLB in the DL kernel.\n\nFinal note to @sheriytm my friend, I tried re-implemented many methods shared by our top participants, but the single most important and effective factor (exclude more complicated techniques mentioned above) seems to be the denoised dataset here, so I share it here as promised to you :) .", "502884": "Well done @ratthachat and thank you very much for generously sharing your findings. I find the problem addressed in this competion fascinating. My goal is to find time later on to build a simple model that can produce nearly as good a result as the top solutions. When and if I do that, I will be posting my updates here.", "502932": "Thank you very much for citing my kernel.", "503069": "Great work, @ratthachat ! Thanks for citing my kernel !", "504122": "ratthachat, I can confirm your observations that denoised data does not predictive performance of tree bsed models. I plugged you data denoised to [my best model](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85146) on the private LB i.e. [my public Catboost kernel](https://www.kaggle.com/sheriytm/5-fold-lstm-atten-fully-commented-original-cpu?scriptVersionId=10248597) and it did not improve the private LB score of 0.62059 but scored the highest of all my models on th public LB with 0.71467. I will try it with my NN model when I resume experimentation on this data." }, "source": "meta" }