I am Grouping Data by customer_ID and Taking the Most Recent Records and was able to score till 0.787
\nSo is their any different method to Group data?
\n Like one notebook i went through was :
\nhttps://www.kaggle.com/code/thedevastator/lag-features-are-all-you-need?rvi=1
\nBut It is difficult to create such features due to Memory overflow.
\nCan someone suggest any other approach?
I am Grouping Data by customer_ID and Taking the Most Recent Records and was able to score till 0.787
\nSo is their any different method to Group data?
\n Like one notebook i went through was :
\nhttps://www.kaggle.com/code/thedevastator/lag-features-are-all-you-need?rvi=1
\nBut It is difficult to create such features due to Memory overflow.
\nCan someone suggest any other approach?
@sarang210 I too tried this approach recently .. no significant results.
", "rawMarkdown": "sarang210 I too tried this approach recently .. no significant results.", "votes": null }, { "id": "1872928", "postDate": "07/27/2022 10:25:19", "content": "I managed to create thedevastator's lag features and not have any overflow issues.
\nCheck out feature_set == 5 in this notebook https://www.kaggle.com/code/mrandri19/0-795-lb-sharing-my-ablations-dart-less-xgb
\nPerformance is not great though.
Thanks @mrandri19 , this one i am already looking, Great snippet by the way.
\nBut again this is lag features approach. I was wondering this is the only way around?
Okey fine, if you got something better please update .
", "rawMarkdown": "Okey fine, if you got something better please update .", "votes": null }, { "id": "1873025", "postDate": "07/27/2022 11:52:08", "content": "There is also this approach: https://www.kaggle.com/code/ragnar123/amex-lgbm-dart-cv-0-7977
\nlast - mean and last - lag1
\nI implemented last - mean in the feature set 7
You might be correct business-wise but in a context of a competition: A quick look at the leaderboard shows that we are all trying to optimize an incredibly tiny number.
\nBy this stage, we might just be fighting on a handful of individual clients.
\nIs it worth it? On the competition: Yes, for sure. There are only a handful of spots in the gold medal zone, there is no way around this.
\nIn real life: That is for AMEX to decide..
From my experience: You are probably correct about the computation load and this type of solutions won't be deployed in real life.
", "rawMarkdown": "You might be correct business-wise but in a context of a competition: A quick look at the leaderboard shows that we are all trying to optimize an incredibly tiny number.\n\nBy this stage, we might just be fighting on a handful of individual clients. \nIs it worth it? On the competition: Yes, for sure. There are only a handful of spots in the gold medal zone, there is no way around this.\nIn real life: That is for AMEX to decide..\n\nFrom my experience: You are probably correct about the computation load and this type of solutions won't be deployed in real life.", "votes": null }, { "id": "1873413", "postDate": "07/27/2022 15:39:25", "content": "Completely agreed.!!
", "rawMarkdown": "Completely agreed.!!", "votes": null }, { "id": "1873414", "postDate": "07/27/2022 15:40:12", "content": "Okey fine i am going to work on this one!
", "rawMarkdown": "Okey fine i am going to work on this one!", "votes": null }, { "id": "1874126", "postDate": "07/28/2022 05:08:57", "content": "I am also stuck on 0.787 with last row per customer.
\nEvery further squeezing is the difference of Kaggle vs business reality where the extra computational costs and maintenance would exceed the marginal gains.
\nIn this competition there seems to be a very high score density, so here it is all about the tiniest findings if you want to rank further up.
", "rawMarkdown": "I am also stuck on 0.787 with last row per customer.\n\nEvery further squeezing is the difference of Kaggle vs business reality where the extra computational costs and maintenance would exceed the marginal gains.\n\nIn this competition there seems to be a very high score density, so here it is all about the tiniest findings if you want to rank further up.", "votes": null }, { "id": "1874134", "postDate": "07/28/2022 05:12:32", "content": "Yeah it is so !!
\nMore features need more computational resources.
\nTrying to optimise as much as i can .
I totally agree with you. Most of our effort on Kaggle is often oriented towards very minor optimization to aggrandize our scores and leaderboard position. We often never even consider these alternatives in real life assignments.
\nI think the art and science of a highly-myopic fine-tuning exercise makes this platform special.
@sarang210 I too tried this approach recently .. no significant results.
", "votes": null, "replies": [ { "id": 1872943, "author_name": "sarang210", "author_url": "", "post_date": "07/27/2022 10:40:31", "content": "Okey fine, if you got something better please update .
", "votes": null, "replies": [] } ] }, { "id": 1872928, "author_name": "mrandri19", "author_url": "", "post_date": "07/27/2022 10:25:19", "content": "I managed to create thedevastator's lag features and not have any overflow issues.
\nCheck out feature_set == 5 in this notebook https://www.kaggle.com/code/mrandri19/0-795-lb-sharing-my-ablations-dart-less-xgb
\nPerformance is not great though.
Thanks @mrandri19 , this one i am already looking, Great snippet by the way.
\nBut again this is lag features approach. I was wondering this is the only way around?
There is also this approach: https://www.kaggle.com/code/ragnar123/amex-lgbm-dart-cv-0-7977
\nlast - mean and last - lag1
\nI implemented last - mean in the feature set 7
Okey fine i am going to work on this one!
", "votes": null, "replies": [] } ] }, { "id": 1873335, "author_name": "thedevastator", "author_url": "", "post_date": "07/27/2022 14:33:06", "content": "You might be correct business-wise but in a context of a competition: A quick look at the leaderboard shows that we are all trying to optimize an incredibly tiny number.
\nBy this stage, we might just be fighting on a handful of individual clients.
\nIs it worth it? On the competition: Yes, for sure. There are only a handful of spots in the gold medal zone, there is no way around this.
\nIn real life: That is for AMEX to decide..
From my experience: You are probably correct about the computation load and this type of solutions won't be deployed in real life.
", "votes": null, "replies": [ { "id": 1873413, "author_name": "sarang210", "author_url": "", "post_date": "07/27/2022 15:39:25", "content": "Completely agreed.!!
", "votes": null, "replies": [] }, { "id": 1874354, "author_name": "ravi20076", "author_url": "", "post_date": "07/28/2022 08:31:05", "content": "I totally agree with you. Most of our effort on Kaggle is often oriented towards very minor optimization to aggrandize our scores and leaderboard position. We often never even consider these alternatives in real life assignments.
\nI think the art and science of a highly-myopic fine-tuning exercise makes this platform special.
I am also stuck on 0.787 with last row per customer.
\nEvery further squeezing is the difference of Kaggle vs business reality where the extra computational costs and maintenance would exceed the marginal gains.
\nIn this competition there seems to be a very high score density, so here it is all about the tiniest findings if you want to rank further up.
", "votes": null, "replies": [ { "id": 1874134, "author_name": "sarang210", "author_url": "", "post_date": "07/28/2022 05:12:32", "content": "Yeah it is so !!
\nMore features need more computational resources.
\nTrying to optimise as much as i can .