{ "id": 80166, "title": "Giving up, making my approach available", "url": "/competitions/vsb-power-line-fault-detection/discussion/80166", "author_name": "Max Halford", "post_date": "2019-02-11T11:03:43.436000", "votes": 86, "comment_count": 42, "views": 0, "content": "

Hey all,

\n\n

I'm giving up this competition as it's clear that RNNs are much better than manual feature extraction + LightGBM etc. However I was quite happy of my approach so I decided to upload it on GitHub. The solution isn't crazy but it might be of interest to you. I implement a neat feature extraction pipeline, and by neat I mean that's it dead simple and extremely fast.

\n\n

Good luck to all of you for the rest of the competition :))

\n\n

EDIT: I had quite a laugh this morning... Thanks everyone for the kind words.

", "messages": [ { "id": 469509, "postDate": "2019-02-11T11:03:43.437Z", "content": "

Hey all,

\n\n

I'm giving up this competition as it's clear that RNNs are much better than manual feature extraction + LightGBM etc. However I was quite happy of my approach so I decided to upload it on GitHub. The solution isn't crazy but it might be of interest to you. I implement a neat feature extraction pipeline, and by neat I mean that's it dead simple and extremely fast.

\n\n

Good luck to all of you for the rest of the competition :))

\n\n

EDIT: I had quite a laugh this morning... Thanks everyone for the kind words.

", "rawMarkdown": "Hey all,\n\nI'm giving up this competition as it's clear that RNNs are much better than manual feature extraction + LightGBM etc. However I was quite happy of my approach so I decided to [upload it on GitHub](https://github.com/MaxHalford/kaggle-vsb-power). The solution isn't crazy but it might be of interest to you. I implement a neat feature extraction pipeline, and by neat I mean that's it dead simple and extremely fast.\n\nGood luck to all of you for the rest of the competition :))\n\nEDIT: I had quite a laugh this morning... Thanks everyone for the kind words.", "votes": 85 }, { "id": 496496, "postDate": "2019-03-22T09:23:02.960Z", "content": "

Well done Max! You made giving up a new winning strategy

", "rawMarkdown": "Well done Max! You made giving up a new winning strategy", "votes": 3, "replies": [ { "id": 496501, "postDate": "2019-03-22T09:27:51.100Z", "content": "

Cheers Antoine :)))

", "rawMarkdown": "Cheers Antoine :)))", "votes": 3 } ] }, { "id": 496153, "postDate": "2019-03-22T00:06:23.930Z", "content": "

Congrats Max, sometimes giving up become best decision :D

", "rawMarkdown": "Congrats Max, sometimes giving up become best decision :D", "votes": 3, "replies": [ { "id": 496668, "postDate": "2019-03-22T13:01:45.033Z", "content": "

I had it all planned (jk)

", "rawMarkdown": "I had it all planned (jk)", "votes": 3 } ] }, { "id": 497349, "postDate": "2019-03-23T12:15:20.507Z", "content": "

I wish there were enough cheater removal to get you gold. But it didn't happen unfortunately. I'm sure we'll see ou in gold soon. Keep up the good sharing spirit you have. And congrats for the result!

", "rawMarkdown": "I wish there were enough cheater removal to get you gold. But it didn't happen unfortunately. I'm sure we'll see ou in gold soon. Keep up the good sharing spirit you have. And congrats for the result!", "votes": 1, "replies": [ { "id": 497614, "postDate": "2019-03-23T18:14:19.927Z", "content": "

Thanks man, it means a lot :). Hope to see you soon!

", "rawMarkdown": "Thanks man, it means a lot :). Hope to see you soon!", "votes": 1 } ] }, { "id": 471644, "postDate": "2019-02-14T17:57:54.570Z", "content": "

Thanks for sharing. It is a pity you don't try RNNs given how well you understand this data.

", "rawMarkdown": "Thanks for sharing. It is a pity you don't try RNNs given how well you understand this data.", "votes": 1, "replies": [ { "id": 471649, "postDate": "2019-02-14T18:03:34.173Z", "content": "

I will one day, but not yet!

", "rawMarkdown": "I will one day, but not yet!", "votes": 1 } ] }, { "id": 471314, "postDate": "2019-02-14T09:48:13.743Z", "content": "

Thanks for sharing and good luck on your PhD and future competitions!

", "rawMarkdown": "Thanks for sharing and good luck on your PhD and future competitions!", "votes": 1 }, { "id": 496212, "postDate": "2019-03-22T01:21:35.237Z", "content": "

I saw your name, and I must say congratulations! I use your feature extraction code for my LGB model and it's in the bronze category. Sadly I didn't trust my own CV.

", "rawMarkdown": "I saw your name, and I must say congratulations! I use your feature extraction code for my LGB model and it's in the bronze category. Sadly I didn't trust my own CV.", "votes": 2, "replies": [ { "id": 496667, "postDate": "2019-03-22T13:00:46.917Z", "content": "

I'm happy it helped! I guess this competition was about doing less, not more. I wonder how many people used my code... It's so funny to me that a 14th place solution was open-sourced and in front of everyone for months!

", "rawMarkdown": "I'm happy it helped! I guess this competition was about doing less, not more. I wonder how many people used my code... It's so funny to me that a 14th place solution was open-sourced and in front of everyone for months!", "votes": 1 } ] }, { "id": 474892, "postDate": "2019-02-20T00:33:43.293Z", "content": "

@MPWARE, It's based on Denoised Signal. I have attached the xgboost feature importance here.

", "rawMarkdown": "@MPWARE, It's based on Denoised Signal. I have attached the xgboost feature importance here.", "votes": 2, "replies": [ { "id": 476242, "postDate": "2019-02-21T21:19:03.143Z", "content": "

Nice!

\n\n

I've tried adversarial validation with the features (min, max, percentiles, range per time buket) + BiLSTM + Attention classifier available in public kernels. Here are some of my results:\n- AUC is quite high (> 0.7) \n- I got similar AUC with same features but with a LightGBM classifier\nFrom what I read in different papers, best would be AUC around 0.5 to have a model not able to distinguish between train and test sample.

\n\n

I've selected train samples similar to test samples by keeping probabilities above 0.75 (prob=1.0 means close to test sample):\n- Adversarial measurements for validation: 288\n- Adversarial ratio for validation: 9.92% (288/2904)\n- Validation labels: Positive: 71, Negative: 793, Ratio: 8.22% (above the 6% for full train dataset)\n- Local: 0.608, LB: 0.607. It seems to work as expected.

\n\n

But problem with adversarial validation is that we cannot use CV anymore and globally I got better scores with CV than with adversarial validation. Anyway, it should help now to evaluate new features and I think it could be used in stratification to balance similar/non similar samples.

\n\n

BTW: Did you try teacher/student network or any other semi-supervised network?

", "rawMarkdown": "Nice!\n\nI've tried adversarial validation with the features (min, max, percentiles, range per time buket) + BiLSTM + Attention classifier available in public kernels. Here are some of my results:\n- AUC is quite high (> 0.7) \n- I got similar AUC with same features but with a LightGBM classifier\nFrom what I read in different papers, best would be AUC around 0.5 to have a model not able to distinguish between train and test sample.\n\nI've selected train samples similar to test samples by keeping probabilities above 0.75 (prob=1.0 means close to test sample):\n- Adversarial measurements for validation: 288\n- Adversarial ratio for validation: 9.92% (288/2904)\n- Validation labels: Positive: 71, Negative: 793, Ratio: 8.22% (above the 6% for full train dataset)\n- Local: 0.608, LB: 0.607. It seems to work as expected.\n\nBut problem with adversarial validation is that we cannot use CV anymore and globally I got better scores with CV than with adversarial validation. Anyway, it should help now to evaluate new features and I think it could be used in stratification to balance similar/non similar samples.\n\nBTW: Did you try teacher/student network or any other semi-supervised network?\n", "votes": 4 }, { "id": 476282, "postDate": "2019-02-21T22:49:15.500Z", "content": "

Thanks for the insights. I didn't try any semi supervised method yet.

", "rawMarkdown": "Thanks for the insights. I didn't try any semi supervised method yet." } ] }, { "id": 474229, "postDate": "2019-02-19T05:21:15.260Z", "content": "

Thanks Max for Sharing. You can give tsfresh package a shot to get more features. \nCheck this :\nhttps://tsfresh.readthedocs.io/en/latest/text/forecasting.html

", "rawMarkdown": "Thanks Max for Sharing. You can give tsfresh package a shot to get more features. \nCheck this :\nhttps://tsfresh.readthedocs.io/en/latest/text/forecasting.html", "votes": 2, "replies": [ { "id": 474536, "postDate": "2019-02-19T14:19:40.263Z", "content": "

Any luck with this package ? I tried many features from this package that I thought would be useful, like number of peaks, entropy, symetry_looking... but it seems that at least with my model they were over-fitting the training set.

", "rawMarkdown": "Any luck with this package ? I tried many features from this package that I thought would be useful, like number of peaks, entropy, symetry_looking... but it seems that at least with my model they were over-fitting the training set.\n", "votes": 1 }, { "id": 474670, "postDate": "2019-02-19T17:23:36.650Z", "content": "

Antoine, Few features which helped me were complexity, non-linearity, entropy and energy decomposition. Using a lot of features were causing overfitting in my case as well , but then i took help of adversarial validation to remove features. I was able to reduce auc (for separation of train/test) of adversarial validation to .63 from 0.95, which helped me to relate my cv and lb score, but since LSTM started to give better result, i stopped using decision tree stuff.

", "rawMarkdown": "Antoine, Few features which helped me were complexity, non-linearity, entropy and energy decomposition. Using a lot of features were causing overfitting in my case as well , but then i took help of adversarial validation to remove features. I was able to reduce auc (for separation of train/test) of adversarial validation to .63 from 0.95, which helped me to relate my cv and lb score, but since LSTM started to give better result, i stopped using decision tree stuff.", "votes": 3 }, { "id": 474829, "postDate": "2019-02-19T21:20:15.110Z", "content": "

Thanks ! I was actually trying to add some of those feature inside my LSTM GRU Model but without any success so far...

", "rawMarkdown": "Thanks ! I was actually trying to add some of those feature inside my LSTM GRU Model but without any success so far..." }, { "id": 474839, "postDate": "2019-02-19T21:44:48.827Z", "content": "

You are adding features after dealing with the randomness ? Without fixing that, i believe, we can't say whether the added features have helped us or not, since the variation because of randomness is quite high.

", "rawMarkdown": "You are adding features after dealing with the randomness ? Without fixing that, i believe, we can't say whether the added features have helped us or not, since the variation because of randomness is quite high. " }, { "id": 474851, "postDate": "2019-02-19T22:14:07.590Z", "content": "

I deal with the randomness by running multiple times the same fold and then taking the average predictions

", "rawMarkdown": "I deal with the randomness by running multiple times the same fold and then taking the average predictions", "votes": 1 }, { "id": 474860, "postDate": "2019-02-19T22:38:48.593Z", "content": "

@HarshitMehta, for non-linearity, it's on raw signal, not with denoised one, correct?

", "rawMarkdown": "@HarshitMehta, for non-linearity, it's on raw signal, not with denoised one, correct?" }, { "id": 492955, "postDate": "2019-03-18T03:08:49.490Z", "content": "

@HarshitMehta, \n'' I was able to reduce auc (for separation of train/test) of adversarial validation to .63 from 0.95,''\nafter that ,the auc of training (only use train data) is still around 0.65?

", "rawMarkdown": "@HarshitMehta, \n'' I was able to reduce auc (for separation of train/test) of adversarial validation to .63 from 0.95,''\nafter that ,the auc of training (only use train data) is still around 0.65?\n" } ] }, { "id": 471288, "postDate": "2019-02-14T08:49:41.373Z", "content": "

Thanks for sharing. \nMax, why giving up ?. still there is time and fight left.

", "rawMarkdown": "Thanks for sharing. \nMax, why giving up ?. still there is time and fight left. ", "votes": 2, "replies": [ { "id": 471298, "postDate": "2019-02-14T09:16:59.953Z", "content": "

I'm just more interested by other competitions and I have a PhD to move forward!

", "rawMarkdown": "I'm just more interested by other competitions and I have a PhD to move forward!", "votes": 4 }, { "id": 496515, "postDate": "2019-03-22T09:40:34.540Z", "content": "

when you gave up and yet you won silver medal! It's almost gold medal if you push harder. Congratulations on your silver medal! :-)

", "rawMarkdown": "when you gave up and yet you won silver medal! It's almost gold medal if you push harder. Congratulations on your silver medal! :-)", "votes": 14 }, { "id": 496569, "postDate": "2019-03-22T10:51:14.143Z", "content": "

I just looked and I actually have a potential gold medal submission... It just didn't get picked automatically by Kaggle.

", "rawMarkdown": "I just looked and I actually have a potential gold medal submission... It just didn't get picked automatically by Kaggle.", "votes": 2 }, { "id": 499814, "postDate": "2019-03-25T08:25:13.593Z", "content": "

I am interested about the difference approach between \"gold medal submission\" and current github repository, thanks for your sharing!

", "rawMarkdown": "I am interested about the difference approach between \"gold medal submission\" and current github repository, thanks for your sharing!" }, { "id": 499817, "postDate": "2019-03-25T08:30:44.380Z", "content": "

@JengHung it's simply a blend with one of the good kernels that was published towards the start of the competition, but I can't remember which.

", "rawMarkdown": "@JengHung it's simply a blend with one of the good kernels that was published towards the start of the competition, but I can't remember which." } ] }, { "id": 503785, "postDate": "2019-03-30T15:07:06.070Z", "content": "

Congratulations Max! We don't want you giving up in the future, because we don't want to miss any good solution!!

", "rawMarkdown": "Congratulations Max! We don't want you giving up in the future, because we don't want to miss any good solution!!" }, { "id": 486764, "postDate": "2019-03-09T11:31:11.840Z", "content": "

Well, your position is far better than mine, so the most reasonable thing for me should be giving up too. However, I decided to continue and maybe I will use your material on Git Hub about which I say thank you in advance. Anyway, different people join the competition with different purposes: mine is learning and reinventing myself and this competition is me helping a lot in order to do that. Have a nice weekend and a nice life, bye!

", "rawMarkdown": "Well, your position is far better than mine, so the most reasonable thing for me should be giving up too. However, I decided to continue and maybe I will use your material on Git Hub about which I say thank you in advance. Anyway, different people join the competition with different purposes: mine is learning and reinventing myself and this competition is me helping a lot in order to do that. Have a nice weekend and a nice life, bye!" }, { "id": 481185, "postDate": "2019-03-01T06:04:58.277Z", "content": "

Thanks for sharing, really helpful. Good luck on your PhD.

", "rawMarkdown": "Thanks for sharing, really helpful. Good luck on your PhD." }, { "id": 470507, "postDate": "2019-02-13T04:59:23.457Z", "rawMarkdown": "", "isDeleted": true }, { "id": 496155, "postDate": "2019-03-22T00:08:42.837Z", "content": "

congras max, and thank you!

", "rawMarkdown": "congras max, and thank you!\n", "votes": 1 }, { "id": 469774, "postDate": "2019-02-11T19:52:15.563Z", "content": "

Thanks for sharing Max.

", "rawMarkdown": "Thanks for sharing Max.", "votes": 1 }, { "id": 492607, "postDate": "2019-03-17T14:00:06.510Z", "content": "

Thanks Max!!

", "rawMarkdown": "Thanks Max!!" }, { "id": 484948, "postDate": "2019-03-06T17:14:54.693Z", "content": "

Thanks for sharing!

", "rawMarkdown": "Thanks for sharing!" }, { "id": 479732, "postDate": "2019-02-27T11:27:50.327Z", "content": "

Thanks for sharing!

", "rawMarkdown": "Thanks for sharing!" }, { "id": 471948, "postDate": "2019-02-15T06:13:07.223Z", "content": "

Thanks for sharing and nice code.

", "rawMarkdown": "Thanks for sharing and nice code." }, { "id": 471487, "postDate": "2019-02-14T14:27:19.453Z", "content": "

Thanks for sharing!

", "rawMarkdown": "Thanks for sharing!" }, { "id": 471178, "postDate": "2019-02-14T05:51:49.330Z", "content": "

Thank for sharing

", "rawMarkdown": "Thank for sharing" }, { "id": 470516, "postDate": "2019-02-13T05:33:11.513Z", "content": "

Thank you for your sharing!

", "rawMarkdown": "Thank you for your sharing!" }, { "id": 469761, "postDate": "2019-02-11T19:30:41.667Z", "content": "

Very nice! Thank you for sharing.

", "rawMarkdown": "Very nice! Thank you for sharing." } ], "comments": [ { "id": 496496, "author_name": "Antoine Plissonneau", "author_url": "", "post_date": "2019-03-22T09:23:02.960000", "content": "

Well done Max! You made giving up a new winning strategy

", "votes": 3, "replies": [ { "id": 496501, "author_name": "Max Halford", "author_url": "", "post_date": "2019-03-22T09:27:51.100000", "content": "

Cheers Antoine :)))

", "votes": 3, "replies": [] } ] }, { "id": 496153, "author_name": "Murat Korkmaz", "author_url": "", "post_date": "2019-03-22T00:06:23.930000", "content": "

Congrats Max, sometimes giving up become best decision :D

", "votes": 3, "replies": [ { "id": 496668, "author_name": "Max Halford", "author_url": "", "post_date": "2019-03-22T13:01:45.033000", "content": "

I had it all planned (jk)

", "votes": 3, "replies": [] } ] }, { "id": 497349, "author_name": "CPMP", "author_url": "", "post_date": "2019-03-23T12:15:20.507000", "content": "

I wish there were enough cheater removal to get you gold. But it didn't happen unfortunately. I'm sure we'll see ou in gold soon. Keep up the good sharing spirit you have. And congrats for the result!

", "votes": 1, "replies": [ { "id": 497614, "author_name": "Max Halford", "author_url": "", "post_date": "2019-03-23T18:14:19.927000", "content": "

Thanks man, it means a lot :). Hope to see you soon!

", "votes": 1, "replies": [] } ] }, { "id": 471644, "author_name": "CPMP", "author_url": "", "post_date": "2019-02-14T17:57:54.570000", "content": "

Thanks for sharing. It is a pity you don't try RNNs given how well you understand this data.

", "votes": 1, "replies": [ { "id": 471649, "author_name": "Max Halford", "author_url": "", "post_date": "2019-02-14T18:03:34.173000", "content": "

I will one day, but not yet!

", "votes": 1, "replies": [] } ] }, { "id": 471314, "author_name": "Daniel Militao", "author_url": "", "post_date": "2019-02-14T09:48:13.743000", "content": "

Thanks for sharing and good luck on your PhD and future competitions!

", "votes": 1, "replies": [] }, { "id": 496212, "author_name": "yukiya", "author_url": "", "post_date": "2019-03-22T01:21:35.237000", "content": "

I saw your name, and I must say congratulations! I use your feature extraction code for my LGB model and it's in the bronze category. Sadly I didn't trust my own CV.

", "votes": 2, "replies": [ { "id": 496667, "author_name": "Max Halford", "author_url": "", "post_date": "2019-03-22T13:00:46.917000", "content": "

I'm happy it helped! I guess this competition was about doing less, not more. I wonder how many people used my code... It's so funny to me that a 14th place solution was open-sourced and in front of everyone for months!

", "votes": 1, "replies": [] } ] }, { "id": 474892, "author_name": "HarshitMehta", "author_url": "", "post_date": "2019-02-20T00:33:43.293000", "content": "

@MPWARE, It's based on Denoised Signal. I have attached the xgboost feature importance here.

", "votes": 2, "replies": [ { "id": 476242, "author_name": "MPWARE", "author_url": "", "post_date": "2019-02-21T21:19:03.143000", "content": "

Nice!

\n\n

I've tried adversarial validation with the features (min, max, percentiles, range per time buket) + BiLSTM + Attention classifier available in public kernels. Here are some of my results:\n- AUC is quite high (> 0.7) \n- I got similar AUC with same features but with a LightGBM classifier\nFrom what I read in different papers, best would be AUC around 0.5 to have a model not able to distinguish between train and test sample.

\n\n

I've selected train samples similar to test samples by keeping probabilities above 0.75 (prob=1.0 means close to test sample):\n- Adversarial measurements for validation: 288\n- Adversarial ratio for validation: 9.92% (288/2904)\n- Validation labels: Positive: 71, Negative: 793, Ratio: 8.22% (above the 6% for full train dataset)\n- Local: 0.608, LB: 0.607. It seems to work as expected.

\n\n

But problem with adversarial validation is that we cannot use CV anymore and globally I got better scores with CV than with adversarial validation. Anyway, it should help now to evaluate new features and I think it could be used in stratification to balance similar/non similar samples.

\n\n

BTW: Did you try teacher/student network or any other semi-supervised network?

", "votes": 4, "replies": [] }, { "id": 476282, "author_name": "HarshitMehta", "author_url": "", "post_date": "2019-02-21T22:49:15.500000", "content": "

Thanks for the insights. I didn't try any semi supervised method yet.

", "votes": 0, "replies": [] } ] }, { "id": 474229, "author_name": "HarshitMehta", "author_url": "", "post_date": "2019-02-19T05:21:15.260000", "content": "

Thanks Max for Sharing. You can give tsfresh package a shot to get more features. \nCheck this :\nhttps://tsfresh.readthedocs.io/en/latest/text/forecasting.html

", "votes": 2, "replies": [ { "id": 474536, "author_name": "Antoine", "author_url": "", "post_date": "2019-02-19T14:19:40.263000", "content": "

Any luck with this package ? I tried many features from this package that I thought would be useful, like number of peaks, entropy, symetry_looking... but it seems that at least with my model they were over-fitting the training set.

", "votes": 1, "replies": [] }, { "id": 474670, "author_name": "HarshitMehta", "author_url": "", "post_date": "2019-02-19T17:23:36.650000", "content": "

Antoine, Few features which helped me were complexity, non-linearity, entropy and energy decomposition. Using a lot of features were causing overfitting in my case as well , but then i took help of adversarial validation to remove features. I was able to reduce auc (for separation of train/test) of adversarial validation to .63 from 0.95, which helped me to relate my cv and lb score, but since LSTM started to give better result, i stopped using decision tree stuff.

", "votes": 3, "replies": [] }, { "id": 474829, "author_name": "Antoine", "author_url": "", "post_date": "2019-02-19T21:20:15.110000", "content": "

Thanks ! I was actually trying to add some of those feature inside my LSTM GRU Model but without any success so far...

", "votes": 0, "replies": [] }, { "id": 474839, "author_name": "HarshitMehta", "author_url": "", "post_date": "2019-02-19T21:44:48.827000", "content": "

You are adding features after dealing with the randomness ? Without fixing that, i believe, we can't say whether the added features have helped us or not, since the variation because of randomness is quite high.

", "votes": 0, "replies": [] }, { "id": 474851, "author_name": "Antoine", "author_url": "", "post_date": "2019-02-19T22:14:07.590000", "content": "

I deal with the randomness by running multiple times the same fold and then taking the average predictions

", "votes": 1, "replies": [] }, { "id": 474860, "author_name": "MPWARE", "author_url": "", "post_date": "2019-02-19T22:38:48.593000", "content": "

@HarshitMehta, for non-linearity, it's on raw signal, not with denoised one, correct?

", "votes": 0, "replies": [] }, { "id": 492955, "author_name": "yumo", "author_url": "", "post_date": "2019-03-18T03:08:49.490000", "content": "

@HarshitMehta, \n'' I was able to reduce auc (for separation of train/test) of adversarial validation to .63 from 0.95,''\nafter that ,the auc of training (only use train data) is still around 0.65?

", "votes": 0, "replies": [] } ] }, { "id": 471288, "author_name": "Manoj", "author_url": "", "post_date": "2019-02-14T08:49:41.373000", "content": "

Thanks for sharing. \nMax, why giving up ?. still there is time and fight left.

", "votes": 2, "replies": [ { "id": 471298, "author_name": "Max Halford", "author_url": "", "post_date": "2019-02-14T09:16:59.953000", "content": "

I'm just more interested by other competitions and I have a PhD to move forward!

", "votes": 4, "replies": [] }, { "id": 496515, "author_name": "", "author_url": "", "post_date": "2019-03-22T09:40:34.540000", "content": "

when you gave up and yet you won silver medal! It's almost gold medal if you push harder. Congratulations on your silver medal! :-)

", "votes": 14, "replies": [] }, { "id": 496569, "author_name": "Max Halford", "author_url": "", "post_date": "2019-03-22T10:51:14.143000", "content": "

I just looked and I actually have a potential gold medal submission... It just didn't get picked automatically by Kaggle.

", "votes": 2, "replies": [] }, { "id": 499814, "author_name": "JengHung", "author_url": "", "post_date": "2019-03-25T08:25:13.593000", "content": "

I am interested about the difference approach between \"gold medal submission\" and current github repository, thanks for your sharing!

", "votes": 0, "replies": [] }, { "id": 499817, "author_name": "Max Halford", "author_url": "", "post_date": "2019-03-25T08:30:44.380000", "content": "

@JengHung it's simply a blend with one of the good kernels that was published towards the start of the competition, but I can't remember which.

", "votes": 0, "replies": [] } ] }, { "id": 503785, "author_name": "Mohammad Azam Khan", "author_url": "", "post_date": "2019-03-30T15:07:06.070000", "content": "

Congratulations Max! We don't want you giving up in the future, because we don't want to miss any good solution!!

", "votes": 0, "replies": [] }, { "id": 486764, "author_name": "Ludovico Ristori", "author_url": "", "post_date": "2019-03-09T11:31:11.840000", "content": "

Well, your position is far better than mine, so the most reasonable thing for me should be giving up too. However, I decided to continue and maybe I will use your material on Git Hub about which I say thank you in advance. Anyway, different people join the competition with different purposes: mine is learning and reinventing myself and this competition is me helping a lot in order to do that. Have a nice weekend and a nice life, bye!

", "votes": 0, "replies": [] }, { "id": 481185, "author_name": "Surchand Wahengbam", "author_url": "", "post_date": "2019-03-01T06:04:58.277000", "content": "

Thanks for sharing, really helpful. Good luck on your PhD.

", "votes": 0, "replies": [] }, { "id": 470507, "author_name": "", "author_url": "", "post_date": "2019-02-13T04:59:23.457000", "content": "", "votes": 0, "replies": [] }, { "id": 496155, "author_name": "yumo", "author_url": "", "post_date": "2019-03-22T00:08:42.837000", "content": "

congras max, and thank you!

", "votes": 1, "replies": [] }, { "id": 469774, "author_name": "mtodisco10", "author_url": "", "post_date": "2019-02-11T19:52:15.563000", "content": "

Thanks for sharing Max.

", "votes": 1, "replies": [] }, { "id": 492607, "author_name": "Pietro Marinelli", "author_url": "", "post_date": "2019-03-17T14:00:06.510000", "content": "

Thanks Max!!

", "votes": 0, "replies": [] }, { "id": 484948, "author_name": "QianSun", "author_url": "", "post_date": "2019-03-06T17:14:54.693000", "content": "

Thanks for sharing!

", "votes": 0, "replies": [] }, { "id": 479732, "author_name": "PGiN", "author_url": "", "post_date": "2019-02-27T11:27:50.327000", "content": "

Thanks for sharing!

", "votes": 0, "replies": [] }, { "id": 471948, "author_name": "LongYin/杰少", "author_url": "", "post_date": "2019-02-15T06:13:07.223000", "content": "

Thanks for sharing and nice code.

", "votes": 0, "replies": [] }, { "id": 471487, "author_name": "kirikei", "author_url": "", "post_date": "2019-02-14T14:27:19.453000", "content": "

Thanks for sharing!

", "votes": 0, "replies": [] }, { "id": 471178, "author_name": "", "author_url": "", "post_date": "2019-02-14T05:51:49.330000", "content": "

Thank for sharing

", "votes": 0, "replies": [] }, { "id": 470516, "author_name": "feifan", "author_url": "", "post_date": "2019-02-13T05:33:11.513000", "content": "

Thank you for your sharing!

", "votes": 0, "replies": [] }, { "id": 469761, "author_name": "Bruce Cragin", "author_url": "", "post_date": "2019-02-11T19:30:41.667000", "content": "

Very nice! Thank you for sharing.

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "469509": "Hey all,\n\nI'm giving up this competition as it's clear that RNNs are much better than manual feature extraction + LightGBM etc. However I was quite happy of my approach so I decided to [upload it on GitHub](https://github.com/MaxHalford/kaggle-vsb-power). The solution isn't crazy but it might be of interest to you. I implement a neat feature extraction pipeline, and by neat I mean that's it dead simple and extremely fast.\n\nGood luck to all of you for the rest of the competition :))\n\nEDIT: I had quite a laugh this morning... Thanks everyone for the kind words.", "496496": "Well done Max! You made giving up a new winning strategy", "496153": "Congrats Max, sometimes giving up become best decision :D", "497349": "I wish there were enough cheater removal to get you gold. But it didn't happen unfortunately. I'm sure we'll see ou in gold soon. Keep up the good sharing spirit you have. And congrats for the result!", "471644": "Thanks for sharing. It is a pity you don't try RNNs given how well you understand this data.", "471314": "Thanks for sharing and good luck on your PhD and future competitions!", "496212": "I saw your name, and I must say congratulations! I use your feature extraction code for my LGB model and it's in the bronze category. Sadly I didn't trust my own CV.", "474892": "@MPWARE, It's based on Denoised Signal. I have attached the xgboost feature importance here.", "474229": "Thanks Max for Sharing. You can give tsfresh package a shot to get more features. \nCheck this :\nhttps://tsfresh.readthedocs.io/en/latest/text/forecasting.html", "471288": "Thanks for sharing. \nMax, why giving up ?. still there is time and fight left. ", "503785": "Congratulations Max! We don't want you giving up in the future, because we don't want to miss any good solution!!", "486764": "Well, your position is far better than mine, so the most reasonable thing for me should be giving up too. However, I decided to continue and maybe I will use your material on Git Hub about which I say thank you in advance. Anyway, different people join the competition with different purposes: mine is learning and reinventing myself and this competition is me helping a lot in order to do that. Have a nice weekend and a nice life, bye!", "481185": "Thanks for sharing, really helpful. Good luck on your PhD.", "470507": "", "496155": "congras max, and thank you!\n", "469774": "Thanks for sharing Max.", "492607": "Thanks Max!!", "484948": "Thanks for sharing!", "479732": "Thanks for sharing!", "471948": "Thanks for sharing and nice code.", "471487": "Thanks for sharing!", "471178": "Thank for sharing", "470516": "Thank you for your sharing!", "469761": "Very nice! Thank you for sharing." } }