Thank you Raddar @raddar and Martin @ragnar123 for sharing so much and helping so many Kaggle participants. You both made this competition so much more enjoyable!
\nRaddar, your analysis of the data noise and your Kaggle dataset Parquets that you shared allowed thousands of participants to easily load the data, build models, and join the competition. Your clean data allowed everyone to achieve more accurate models.
\nMartin, your LGBM Dart notebook is amazing. So many Kaggler's including myself did not know about Dart, i would never have built such an accurate final model without your guidance. Using Dart gave at least a +0.001 boost to all GBT models!
Thank you Raddar and Martin. Kagglers, please join me in thanking Raddar and Martin for their generous contributions.
\nUPDATE: Thank you @ambrosm, @tilii7, @thedevastator, @aquatic, @roberthatch for your helpful contributions. I read all your discussions and notebooks. All of your contributions helped me and others. I'm adding all links that Kagglers post in the comments below. There was a lot of sharing in this competition. It was all very helpful and wonderful.
", "messages": [ { "id": "1912766", "postDate": "08/25/2022 00:51:11", "content": "Thank you Raddar @raddar and Martin @ragnar123 for sharing so much and helping so many Kaggle participants. You both made this competition so much more enjoyable!
\nRaddar, your analysis of the data noise and your Kaggle dataset Parquets that you shared allowed thousands of participants to easily load the data, build models, and join the competition. Your clean data allowed everyone to achieve more accurate models.
\nMartin, your LGBM Dart notebook is amazing. So many Kaggler's including myself did not know about Dart, i would never have built such an accurate final model without your guidance. Using Dart gave at least a +0.001 boost to all GBT models!
Thank you Raddar and Martin. Kagglers, please join me in thanking Raddar and Martin for their generous contributions.
\nUPDATE: Thank you @ambrosm, @tilii7, @thedevastator, @aquatic, @roberthatch for your helpful contributions. I read all your discussions and notebooks. All of your contributions helped me and others. I'm adding all links that Kagglers post in the comments below. There was a lot of sharing in this competition. It was all very helpful and wonderful.
", "rawMarkdown": "Thank you Raddar @raddar and Martin @ragnar123 for sharing so much and helping so many Kaggle participants. You both made this competition so much more enjoyable!\n\nRaddar, your analysis of the data noise and your Kaggle dataset Parquets that you shared allowed thousands of participants to easily load the data, build models, and join the competition. Your clean data allowed everyone to achieve more accurate models. \n\nMartin, your LGBM Dart notebook is amazing. So many Kaggler's including myself did not know about Dart, i would never have built such an accurate final model without your guidance. Using Dart gave at least a `+0.001` boost to all GBT models!\n\nThank you Raddar and Martin. Kagglers, please join me in thanking Raddar and Martin for their generous contributions.\n\nUPDATE: Thank you @ambrosm, @tilii7, @thedevastator, @aquatic, @roberthatch for your helpful contributions. I read all your discussions and notebooks. All of your contributions helped me and others. I'm adding all links that Kagglers post in the comments below. There was a lot of sharing in this competition. It was all very helpful and wonderful.", "votes": null }, { "id": "1912800", "postDate": "08/25/2022 01:16:11", "content": "Thanks to @raddar and @ragnar123 .. Raddar datasets were the backbone of all great models and Ragnar dart the engine
", "rawMarkdown": "Thanks to @raddar and @ragnar123 .. Raddar datasets were the backbone of all great models and Ragnar dart the engine", "votes": null }, { "id": "1912805", "postDate": "08/25/2022 01:19:52", "content": "From start to finish RADDAR , Martin, @ambrosm and your posts were daily readings for me. Really, this is the first time I started getting a feel for EDA and I have you guys to thank for it. Also @tilii7 , @thedevastator , and @aquatic for all the neat conversations (that I lurked lol). You guys are awesome.
", "rawMarkdown": "From start to finish **RADDAR** , Martin, @ambrosm and your posts were daily readings for me. Really, this is the first time I started getting a feel for EDA and I have you guys to thank for it. Also @tilii7 , @thedevastator , and @aquatic for all the neat conversations (that I lurked lol). You guys are awesome.", "votes": null }, { "id": "1912825", "postDate": "08/25/2022 01:38:50", "content": "Thank you @cdeotte ! With your help I was able to participate at all and get running on my first kaggle competition. Not to mention helping me find @raddar's amazing work! Big thanks to Martin for inspiring a look at what's possible, and @thedevastator for giving constructive criticism and letting me know someone was trying my pyramid work :)
", "rawMarkdown": "Thank you @cdeotte ! With your help I was able to participate at all and get running on my first kaggle competition. Not to mention helping me find @raddar's amazing work! Big thanks to Martin for inspiring a look at what's possible, and @thedevastator for giving constructive criticism and letting me know someone was trying my pyramid work :)", "votes": null }, { "id": "1912828", "postDate": "08/25/2022 01:41:04", "content": "I would not have joined if I didn't see those Parquets dataset, so thank you so much @raddar
\nIt's still dizzying to think how you manage to integerized those values 👀
\nA great baseline means a lot to a competition, so thanks so much @ragnar123 , also to you @cdeotte !
\nI would like to add @ambrosm (what happen to your avatar? ) for the discussion and @tilii7 for the kinky shape t-sne plot 😄
Agreed 100%
\nI was using XG Boost and if I had only used that my final submission using would've probably would've been in the .803-.804 range. Instead, I iterated on @ragnar's notebook, added some of my own FE and with the help of a lot of ensembles, my final score was in the .807s. Additionally, @raddar data was a tremendous help as I really didn't have to worry about memory and could just focus on making my models better. Thanks to you both.
", "rawMarkdown": "Agreed 100% \n\nI was using XG Boost and if I had only used that my final submission using would've probably would've been in the .803-.804 range. Instead, I iterated on @ragnar's notebook, added some of my own FE and with the help of a lot of ensembles, my final score was in the .807s. Additionally, @raddar data was a tremendous help as I really didn't have to worry about memory and could just focus on making my models better. Thanks to you both.", "votes": null }, { "id": "1912848", "postDate": "08/25/2022 02:08:29", "content": "Thank you too, @roberthatch I used your XGBoost pyramid approach on my preprocessed data. It achieved CV 0.7995 on training data. Unfortunately, it underperformed your original notebook on the private test set. I made my notebooks public:
\n\n\nhttps://www.kaggle.com/code/lhagiimn/amex-xgboost-pyramid-cv-0-799
\n
\n https://www.kaggle.com/code/lhagiimn/xgboost-pyramid-inference-cv-0-7995
Thank you, @raddar. You are correct, it was a time-series competition. There is a solution in my submissions to get the silver medal. Although I selected the solutions with the highest CV score, I couldn't achieve a good score on the private test set. I think the reason is probably that the distribution of some input variables was different for the training and test sets.
\nBTW, I always learn a lot from Kaggle and those who make valuable contributions to the competitions. I learned a lot this time too. Thank you all.
", "rawMarkdown": "Thank you too, @roberthatch I used your XGBoost pyramid approach on my preprocessed data. It achieved CV 0.7995 on training data. Unfortunately, it underperformed your original notebook on the private test set. I made my notebooks public:\n> https://www.kaggle.com/code/lhagiimn/amex-xgboost-pyramid-cv-0-799\n> https://www.kaggle.com/code/lhagiimn/xgboost-pyramid-inference-cv-0-7995\n\nThank you, @raddar. You are correct, it was a time-series competition. There is a solution in my submissions to get the silver medal. Although I selected the solutions with the highest CV score, I couldn't achieve a good score on the private test set. I think the reason is probably that the distribution of some input variables was different for the training and test sets.\n\nBTW, I always learn a lot from Kaggle and those who make valuable contributions to the competitions. I learned a lot this time too. Thank you all.", "votes": null }, { "id": "1912863", "postDate": "08/25/2022 02:31:41", "content": "It is my first competition with medals and I learned so much from the codes and discussions generously posted here. It was a great experience. Thank you all!
", "rawMarkdown": "It is my first competition with medals and I learned so much from the codes and discussions generously posted here. It was a great experience. Thank you all!", "votes": null }, { "id": "1912906", "postDate": "08/25/2022 03:12:30", "content": "Deserved recognition to the people that contributed the most.
\nI learned many things and also learned about dart boosting. I couldn't find success with dart though, neither with LGBM or XGB. But I also couldn't dedicate much time on this because default gbtree boosting ran 5 folds in 10 minutes and with dart it took like 4 hours…
", "rawMarkdown": "Deserved recognition to the people that contributed the most.\n\nI learned many things and also learned about dart boosting. I couldn't find success with dart though, neither with LGBM or XGB. But I also couldn't dedicate much time on this because default gbtree boosting ran 5 folds in 10 minutes and with dart it took like 4 hours...", "votes": null }, { "id": "1913587", "postDate": "08/25/2022 11:41:07", "content": "For me @raddar is the moral winner of this competition
", "rawMarkdown": "For me @raddar is the moral winner of this competition", "votes": null }, { "id": "1913879", "postDate": "08/25/2022 14:55:48", "content": "thank you @cdeotte I first mainly used your xgb notebook as I am sure many others have
", "rawMarkdown": "thank you @cdeotte I first mainly used your xgb notebook as I am sure many others have", "votes": null } ], "comments": [ { "id": 1912800, "author_name": "gauravbrills", "author_url": "", "post_date": "08/25/2022 01:16:11", "content": "Thanks to @raddar and @ragnar123 .. Raddar datasets were the backbone of all great models and Ragnar dart the engine
", "votes": null, "replies": [] }, { "id": 1912805, "author_name": "lilgaussy", "author_url": "", "post_date": "08/25/2022 01:19:52", "content": "From start to finish RADDAR , Martin, @ambrosm and your posts were daily readings for me. Really, this is the first time I started getting a feel for EDA and I have you guys to thank for it. Also @tilii7 , @thedevastator , and @aquatic for all the neat conversations (that I lurked lol). You guys are awesome.
", "votes": null, "replies": [] }, { "id": 1912825, "author_name": "roberthatch", "author_url": "", "post_date": "08/25/2022 01:38:50", "content": "Thank you @cdeotte ! With your help I was able to participate at all and get running on my first kaggle competition. Not to mention helping me find @raddar's amazing work! Big thanks to Martin for inspiring a look at what's possible, and @thedevastator for giving constructive criticism and letting me know someone was trying my pyramid work :)
", "votes": null, "replies": [ { "id": 1912848, "author_name": "lhagiimn", "author_url": "", "post_date": "08/25/2022 02:08:29", "content": "Thank you too, @roberthatch I used your XGBoost pyramid approach on my preprocessed data. It achieved CV 0.7995 on training data. Unfortunately, it underperformed your original notebook on the private test set. I made my notebooks public:
\n\n\nhttps://www.kaggle.com/code/lhagiimn/amex-xgboost-pyramid-cv-0-799
\n
\n https://www.kaggle.com/code/lhagiimn/xgboost-pyramid-inference-cv-0-7995
Thank you, @raddar. You are correct, it was a time-series competition. There is a solution in my submissions to get the silver medal. Although I selected the solutions with the highest CV score, I couldn't achieve a good score on the private test set. I think the reason is probably that the distribution of some input variables was different for the training and test sets.
\nBTW, I always learn a lot from Kaggle and those who make valuable contributions to the competitions. I learned a lot this time too. Thank you all.
", "votes": null, "replies": [] } ] }, { "id": 1912828, "author_name": "nyleve", "author_url": "", "post_date": "08/25/2022 01:41:04", "content": "I would not have joined if I didn't see those Parquets dataset, so thank you so much @raddar
\nIt's still dizzying to think how you manage to integerized those values 👀
\nA great baseline means a lot to a competition, so thanks so much @ragnar123 , also to you @cdeotte !
\nI would like to add @ambrosm (what happen to your avatar? ) for the discussion and @tilii7 for the kinky shape t-sne plot 😄
Agreed 100%
\nI was using XG Boost and if I had only used that my final submission using would've probably would've been in the .803-.804 range. Instead, I iterated on @ragnar's notebook, added some of my own FE and with the help of a lot of ensembles, my final score was in the .807s. Additionally, @raddar data was a tremendous help as I really didn't have to worry about memory and could just focus on making my models better. Thanks to you both.
", "votes": null, "replies": [] }, { "id": 1912863, "author_name": "gehallak", "author_url": "", "post_date": "08/25/2022 02:31:41", "content": "It is my first competition with medals and I learned so much from the codes and discussions generously posted here. It was a great experience. Thank you all!
", "votes": null, "replies": [] }, { "id": 1912906, "author_name": "hinepo", "author_url": "", "post_date": "08/25/2022 03:12:30", "content": "Deserved recognition to the people that contributed the most.
\nI learned many things and also learned about dart boosting. I couldn't find success with dart though, neither with LGBM or XGB. But I also couldn't dedicate much time on this because default gbtree boosting ran 5 folds in 10 minutes and with dart it took like 4 hours…
", "votes": null, "replies": [] }, { "id": 1913587, "author_name": "canonicalized", "author_url": "", "post_date": "08/25/2022 11:41:07", "content": "For me @raddar is the moral winner of this competition
", "votes": null, "replies": [] }, { "id": 1913879, "author_name": "pyagoubi", "author_url": "", "post_date": "08/25/2022 14:55:48", "content": "thank you @cdeotte I first mainly used your xgb notebook as I am sure many others have
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1912766": "Thank you Raddar @raddar and Martin @ragnar123 for sharing so much and helping so many Kaggle participants. You both made this competition so much more enjoyable!\n\nRaddar, your analysis of the data noise and your Kaggle dataset Parquets that you shared allowed thousands of participants to easily load the data, build models, and join the competition. Your clean data allowed everyone to achieve more accurate models. \n\nMartin, your LGBM Dart notebook is amazing. So many Kaggler's including myself did not know about Dart, i would never have built such an accurate final model without your guidance. Using Dart gave at least a `+0.001` boost to all GBT models!\n\nThank you Raddar and Martin. Kagglers, please join me in thanking Raddar and Martin for their generous contributions.\n\nUPDATE: Thank you @ambrosm, @tilii7, @thedevastator, @aquatic, @roberthatch for your helpful contributions. I read all your discussions and notebooks. All of your contributions helped me and others. I'm adding all links that Kagglers post in the comments below. There was a lot of sharing in this competition. It was all very helpful and wonderful.", "1912800": "Thanks to @raddar and @ragnar123 .. Raddar datasets were the backbone of all great models and Ragnar dart the engine", "1912805": "From start to finish **RADDAR** , Martin, @ambrosm and your posts were daily readings for me. Really, this is the first time I started getting a feel for EDA and I have you guys to thank for it. Also @tilii7 , @thedevastator , and @aquatic for all the neat conversations (that I lurked lol). You guys are awesome.", "1912825": "Thank you @cdeotte ! With your help I was able to participate at all and get running on my first kaggle competition. Not to mention helping me find @raddar's amazing work! Big thanks to Martin for inspiring a look at what's possible, and @thedevastator for giving constructive criticism and letting me know someone was trying my pyramid work :)", "1912828": "I would not have joined if I didn't see those Parquets dataset, so thank you so much @raddar\nIt's still dizzying to think how you manage to integerized those values 👀\nA great baseline means a lot to a competition, so thanks so much @ragnar123 , also to you @cdeotte ! \nI would like to add @ambrosm (what happen to your avatar? ) for the discussion and @tilii7 for the kinky shape t-sne plot 😄", "1912830": "Agreed 100% \n\nI was using XG Boost and if I had only used that my final submission using would've probably would've been in the .803-.804 range. Instead, I iterated on @ragnar's notebook, added some of my own FE and with the help of a lot of ensembles, my final score was in the .807s. Additionally, @raddar data was a tremendous help as I really didn't have to worry about memory and could just focus on making my models better. Thanks to you both.", "1912848": "Thank you too, @roberthatch I used your XGBoost pyramid approach on my preprocessed data. It achieved CV 0.7995 on training data. Unfortunately, it underperformed your original notebook on the private test set. I made my notebooks public:\n> https://www.kaggle.com/code/lhagiimn/amex-xgboost-pyramid-cv-0-799\n> https://www.kaggle.com/code/lhagiimn/xgboost-pyramid-inference-cv-0-7995\n\nThank you, @raddar. You are correct, it was a time-series competition. There is a solution in my submissions to get the silver medal. Although I selected the solutions with the highest CV score, I couldn't achieve a good score on the private test set. I think the reason is probably that the distribution of some input variables was different for the training and test sets.\n\nBTW, I always learn a lot from Kaggle and those who make valuable contributions to the competitions. I learned a lot this time too. Thank you all.", "1912863": "It is my first competition with medals and I learned so much from the codes and discussions generously posted here. It was a great experience. Thank you all!", "1912906": "Deserved recognition to the people that contributed the most.\n\nI learned many things and also learned about dart boosting. I couldn't find success with dart though, neither with LGBM or XGB. But I also couldn't dedicate much time on this because default gbtree boosting ran 5 folds in 10 minutes and with dart it took like 4 hours...", "1913587": "For me @raddar is the moral winner of this competition", "1913879": "thank you @cdeotte I first mainly used your xgb notebook as I am sure many others have" }, "source": "meta" }