My current single model is CV 3.6360, LB 3.679.
\nI am using 30-200 features, depending on what I want to achieve. (I probably threw away hundreds of features which did not improve CV).
\nI haven't done much ensembling yet, but just throwing in some different LGBMs and NNs gave me my current best score of CV 3.6305, LB 3.675.
\nI am trusting my CV (for now), so I have done nothing to fit the LB (like manual postprocessing).
I am looking for a similar CV-scoring person/team, atleast aiming for a gold-medal.
\n\nI will be mostly available throughout the day (01:00-15:00 UTC), everyday.
\nI hope to learn a lot from you :)
My current single model is CV 3.6360, LB 3.679.
\nI am using 30-200 features, depending on what I want to achieve. (I probably threw away hundreds of features which did not improve CV).
\nI haven't done much ensembling yet, but just throwing in some different LGBMs and NNs gave me my current best score of CV 3.6305, LB 3.675.
\nI am trusting my CV (for now), so I have done nothing to fit the LB (like manual postprocessing).
I am looking for a similar CV-scoring person/team, atleast aiming for a gold-medal.
\n\nI will be mostly available throughout the day (01:00-15:00 UTC), everyday.
\nI hope to learn a lot from you :)
My current single model is CV 3.638+, I think it might be overfitting ...
", "rawMarkdown": "My current single model is CV 3.638+, I think it might be overfitting ...", "votes": null }, { "id": "465446", "postDate": "02/03/2019 05:50:25", "content": "What makes you think so?
\nWith a perfect 0.5 score of adversarial validation, I would lean towards trusting CV (unless the host cherry picked the test data by the target)\nI haven't looked too much into the train/test difference, so take it with grain of salt though
Hi @pocket!\nCurrently I'm on a blended, no post processing score of 3.634 cv and 3.679 LB. (which is very similar to your single model)\nI admit i have a weak single model. \n3.646 cv and 3.695 LB.\nBut it is simple, only 36 features. \nThe main thing I've worked on and still working on is a strong and interesting blending pipeline that i've used to go down from LB 3.695 to 3.679
\n\nI also trust my CV, it remains consistent correlation with LB.\nLet me know if you wanna Skype to discuss things. You can pm me your Skype handle.
\n\nI'm pretty sure that if we implement your FE and single model into the blending pipeline i developed - we can make a significant jump.\nI got 2 silver medals in last couple of competitions, and i aim for the gold =)
", "rawMarkdown": "Hi @pocket!\nCurrently I'm on a blended, no post processing score of 3.634 cv and 3.679 LB. (which is very similar to your single model)\nI admit i have a weak single model. \n3.646 cv and 3.695 LB.\nBut it is simple, only 36 features. \nThe main thing I've worked on and still working on is a strong and interesting blending pipeline that i've used to go down from LB 3.695 to 3.679\n\nI also trust my CV, it remains consistent correlation with LB.\nLet me know if you wanna Skype to discuss things. You can pm me your Skype handle.\n\nI'm pretty sure that if we implement your FE and single model into the blending pipeline i developed - we can make a significant jump.\nI got 2 silver medals in last couple of competitions, and i aim for the gold =)", "votes": null }, { "id": "465507", "postDate": "02/03/2019 09:46:24", "content": "How about a team up? Our current model's result on LB is 3.680, CV is 3.650, no manual post processing. Currently we are following the 3 model approach. After regression of non outlier model, classification of outliers and merging with our best model which is 3.680 is giving our current result on LB. We are not currently using the manual post processing coz of the fear of over fitting to LB.
", "rawMarkdown": "How about a team up? Our current model's result on LB is 3.680, CV is 3.650, no manual post processing. Currently we are following the 3 model approach. After regression of non outlier model, classification of outliers and merging with our best model which is **3.680** is giving our current result on LB. We are not currently using the manual post processing coz of the fear of over fitting to LB.", "votes": null }, { "id": "465585", "postDate": "02/03/2019 14:29:40", "content": "Hi finetuned.
\nThank you for your offer, but I don't think we can reach gold, at least for now.
Hi AmirH, you do indeed have a interesting pipeline (after reading your talkingdata solution)
\nI think there is a long way to gold, and I am not sure if we can reach it with just my model+blend
Hi @pocket,\nThanks for your response.\nThere's still a long way to go until end of competition =)\nI surely don't intend to simply plug your model and say we're done =D
\n\nI think the way to go is for us to share our features and approaches and implement them together.\nso i intend to use your features and ideas of your single model to further improve my current models, and also share with you some of my approaches so we can work on additional models for diversity + utilize the blending pipeline.
\n\nThere's still much to be done and many things to explore.
", "rawMarkdown": "Hi @pocket,\nThanks for your response.\nThere's still a long way to go until end of competition =)\nI surely don't intend to simply plug your model and say we're done =D\n\nI think the way to go is for us to share our features and approaches and implement them together.\nso i intend to use your features and ideas of your single model to further improve my current models, and also share with you some of my approaches so we can work on additional models for diversity + utilize the blending pipeline.\n\nThere's still much to be done and many things to explore.", "votes": null }, { "id": "465610", "postDate": "02/03/2019 15:19:42", "content": "We have two positions, are you still available?
", "rawMarkdown": "We have two positions, are you still available?", "votes": null }, { "id": "465662", "postDate": "02/03/2019 18:02:15", "content": "I have my single LGB performing 3.78. After some basic staking , its 3.78. My current score is the post processing stuff,\nLet me know if you r interested.
", "rawMarkdown": "I have my single LGB performing 3.78. After some basic staking , its 3.78. My current score is the post processing stuff,\nLet me know if you r interested.", "votes": null }, { "id": "465780", "postDate": "02/04/2019 01:54:57", "content": "I have sent you an e-mail message to you kaggle account.
", "rawMarkdown": "I have sent you an e-mail message to you kaggle account.", "votes": null }, { "id": "465781", "postDate": "02/04/2019 01:55:20", "content": "I think you have mistaken your numbers :)
", "rawMarkdown": "I think you have mistaken your numbers :)", "votes": null }, { "id": "465857", "postDate": "02/04/2019 07:10:32", "content": "AmirH, Thank you for showing your dedication :)\nI decided to merge with another team, maybe we can team next time.\nWill be interested in your solution in the end :D
", "rawMarkdown": "AmirH, Thank you for showing your dedication :)\nI decided to merge with another team, maybe we can team next time.\nWill be interested in your solution in the end :D", "votes": null }, { "id": "465950", "postDate": "02/04/2019 11:33:57", "content": "Sure thing @pocket.\nDon't overfit to LB!
", "rawMarkdown": "Sure thing @pocket.\nDon't overfit to LB!", "votes": null }, { "id": "466411", "postDate": "02/05/2019 10:49:01", "content": "More details please
", "rawMarkdown": "More details please", "votes": null }, { "id": "466949", "postDate": "02/06/2019 08:49:35", "content": "Hi, I'm new in kaggle. What is the manual post processing?\nI would be grateful if you could answer me. :)
", "rawMarkdown": "Hi, I'm new in kaggle. What is the manual post processing?\nI would be grateful if you could answer me. :)", "votes": null } ], "comments": [ { "id": 465442, "author_name": "lmdeliangmi", "author_url": "", "post_date": "02/03/2019 05:37:41", "content": "My current single model is CV 3.638+, I think it might be overfitting ...
", "votes": null, "replies": [ { "id": 465446, "author_name": "pocketsuteado", "author_url": "", "post_date": "02/03/2019 05:50:25", "content": "What makes you think so?
\nWith a perfect 0.5 score of adversarial validation, I would lean towards trusting CV (unless the host cherry picked the test data by the target)\nI haven't looked too much into the train/test difference, so take it with grain of salt though
Hi @pocket!\nCurrently I'm on a blended, no post processing score of 3.634 cv and 3.679 LB. (which is very similar to your single model)\nI admit i have a weak single model. \n3.646 cv and 3.695 LB.\nBut it is simple, only 36 features. \nThe main thing I've worked on and still working on is a strong and interesting blending pipeline that i've used to go down from LB 3.695 to 3.679
\n\nI also trust my CV, it remains consistent correlation with LB.\nLet me know if you wanna Skype to discuss things. You can pm me your Skype handle.
\n\nI'm pretty sure that if we implement your FE and single model into the blending pipeline i developed - we can make a significant jump.\nI got 2 silver medals in last couple of competitions, and i aim for the gold =)
", "votes": null, "replies": [ { "id": 465589, "author_name": "pocketsuteado", "author_url": "", "post_date": "02/03/2019 14:39:57", "content": "Hi AmirH, you do indeed have a interesting pipeline (after reading your talkingdata solution)
\nI think there is a long way to gold, and I am not sure if we can reach it with just my model+blend
Hi @pocket,\nThanks for your response.\nThere's still a long way to go until end of competition =)\nI surely don't intend to simply plug your model and say we're done =D
\n\nI think the way to go is for us to share our features and approaches and implement them together.\nso i intend to use your features and ideas of your single model to further improve my current models, and also share with you some of my approaches so we can work on additional models for diversity + utilize the blending pipeline.
\n\nThere's still much to be done and many things to explore.
", "votes": null, "replies": [] }, { "id": 465857, "author_name": "pocketsuteado", "author_url": "", "post_date": "02/04/2019 07:10:32", "content": "AmirH, Thank you for showing your dedication :)\nI decided to merge with another team, maybe we can team next time.\nWill be interested in your solution in the end :D
", "votes": null, "replies": [] }, { "id": 465950, "author_name": "tpthegreat", "author_url": "", "post_date": "02/04/2019 11:33:57", "content": "Sure thing @pocket.\nDon't overfit to LB!
", "votes": null, "replies": [] } ] }, { "id": 465507, "author_name": "manu845", "author_url": "", "post_date": "02/03/2019 09:46:24", "content": "How about a team up? Our current model's result on LB is 3.680, CV is 3.650, no manual post processing. Currently we are following the 3 model approach. After regression of non outlier model, classification of outliers and merging with our best model which is 3.680 is giving our current result on LB. We are not currently using the manual post processing coz of the fear of over fitting to LB.
", "votes": null, "replies": [ { "id": 465585, "author_name": "pocketsuteado", "author_url": "", "post_date": "02/03/2019 14:29:40", "content": "Hi finetuned.
\nThank you for your offer, but I don't think we can reach gold, at least for now.
Hi, I'm new in kaggle. What is the manual post processing?\nI would be grateful if you could answer me. :)
", "votes": null, "replies": [] } ] }, { "id": 465610, "author_name": "marcuslin", "author_url": "", "post_date": "02/03/2019 15:19:42", "content": "We have two positions, are you still available?
", "votes": null, "replies": [ { "id": 465780, "author_name": "pocketsuteado", "author_url": "", "post_date": "02/04/2019 01:54:57", "content": "I have sent you an e-mail message to you kaggle account.
", "votes": null, "replies": [] } ] }, { "id": 465662, "author_name": "subikashpal", "author_url": "", "post_date": "02/03/2019 18:02:15", "content": "I have my single LGB performing 3.78. After some basic staking , its 3.78. My current score is the post processing stuff,\nLet me know if you r interested.
", "votes": null, "replies": [ { "id": 465781, "author_name": "pocketsuteado", "author_url": "", "post_date": "02/04/2019 01:55:20", "content": "I think you have mistaken your numbers :)
", "votes": null, "replies": [] } ] }, { "id": 466411, "author_name": "chittaranjan021", "author_url": "", "post_date": "02/05/2019 10:49:01", "content": "More details please
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "465440": "My current single model is CV 3.6360, LB 3.679. \nI am using 30-200 features, depending on what I want to achieve. (I probably threw away hundreds of features which did not improve CV). \nI haven't done much ensembling yet, but just throwing in some different LGBMs and NNs gave me my current best score of CV 3.6305, LB 3.675. \nI am trusting my CV (for now), so I have done nothing to fit the LB (like manual postprocessing). \n\nI am looking for a similar CV-scoring person/team, atleast aiming for a gold-medal.\n\nI will be mostly available throughout the day (01:00-15:00 UTC), everyday. \nI hope to learn a lot from you :)", "465442": "My current single model is CV 3.638+, I think it might be overfitting ...", "465446": "What makes you think so? \nWith a perfect 0.5 score of adversarial validation, I would lean towards trusting CV (unless the host cherry picked the test data by the target)\nI haven't looked too much into the train/test difference, so take it with grain of salt though", "465459": "Hi @pocket!\nCurrently I'm on a blended, no post processing score of 3.634 cv and 3.679 LB. (which is very similar to your single model)\nI admit i have a weak single model. \n3.646 cv and 3.695 LB.\nBut it is simple, only 36 features. \nThe main thing I've worked on and still working on is a strong and interesting blending pipeline that i've used to go down from LB 3.695 to 3.679\n\nI also trust my CV, it remains consistent correlation with LB.\nLet me know if you wanna Skype to discuss things. You can pm me your Skype handle.\n\nI'm pretty sure that if we implement your FE and single model into the blending pipeline i developed - we can make a significant jump.\nI got 2 silver medals in last couple of competitions, and i aim for the gold =)", "465507": "How about a team up? Our current model's result on LB is 3.680, CV is 3.650, no manual post processing. Currently we are following the 3 model approach. After regression of non outlier model, classification of outliers and merging with our best model which is **3.680** is giving our current result on LB. We are not currently using the manual post processing coz of the fear of over fitting to LB.", "465585": "Hi finetuned. \nThank you for your offer, but I don't think we can reach gold, at least for now.", "465589": "Hi AmirH, you do indeed have a interesting pipeline (after reading your talkingdata solution) \nI think there is a long way to gold, and I am not sure if we can reach it with just my model+blend", "465596": "Hi @pocket,\nThanks for your response.\nThere's still a long way to go until end of competition =)\nI surely don't intend to simply plug your model and say we're done =D\n\nI think the way to go is for us to share our features and approaches and implement them together.\nso i intend to use your features and ideas of your single model to further improve my current models, and also share with you some of my approaches so we can work on additional models for diversity + utilize the blending pipeline.\n\nThere's still much to be done and many things to explore.", "465610": "We have two positions, are you still available?", "465662": "I have my single LGB performing 3.78. After some basic staking , its 3.78. My current score is the post processing stuff,\nLet me know if you r interested.", "465780": "I have sent you an e-mail message to you kaggle account.", "465781": "I think you have mistaken your numbers :)", "465857": "AmirH, Thank you for showing your dedication :)\nI decided to merge with another team, maybe we can team next time.\nWill be interested in your solution in the end :D", "465950": "Sure thing @pocket.\nDon't overfit to LB!", "466411": "More details please", "466949": "Hi, I'm new in kaggle. What is the manual post processing?\nI would be grateful if you could answer me. :)" }, "source": "meta" }