My current method contains three steps:
\nThe recall rate of the result in the first step is about 9% (which is really low). However, when I tried to generate more candidates to improve the overall recall rate, the performance of my lgb ranker dramatically decreased, so as the CV score on the valid set😨. Downsampling negative samples couldn't solve the problem.
\nI'm new to this field, did I miss some essential steps in the process? Or do you guys have any suggestions? Thanks a lot!!!
\nBTW, I'm looking for teammates who have roughly the same or higher score as me (currently 0.0255) and are willing to spend time on this competition (I can spend around 6 hours each day)😀
\nUpdate: I missed the discussion here, will try the mentioned method and update the result here.
\nUpdate2: Seems like the problem still exists
", "messages": [ { "id": "1771078", "postDate": "04/28/2022 21:38:34", "content": "My current method contains three steps:
\nThe recall rate of the result in the first step is about 9% (which is really low). However, when I tried to generate more candidates to improve the overall recall rate, the performance of my lgb ranker dramatically decreased, so as the CV score on the valid set😨. Downsampling negative samples couldn't solve the problem.
\nI'm new to this field, did I miss some essential steps in the process? Or do you guys have any suggestions? Thanks a lot!!!
\nBTW, I'm looking for teammates who have roughly the same or higher score as me (currently 0.0255) and are willing to spend time on this competition (I can spend around 6 hours each day)😀
\nUpdate: I missed the discussion here, will try the mentioned method and update the result here.
\nUpdate2: Seems like the problem still exists
", "rawMarkdown": "My current method contains three steps:\n\n1. retrieve/recall items by some basic rules (buying history, popularity, etc)\n2. label items from step 1 according to the customer's behavior in the valid set\n3. train an lgb ranker to rank the candidates.\n\nThe recall rate of the result in the first step is about 9% (which is really low). However, when I tried to generate more candidates to improve the overall recall rate, the performance of my lgb ranker dramatically decreased, so as the CV score on the valid set😨. Downsampling negative samples couldn't solve the problem.\n\nI'm new to this field, did I miss some essential steps in the process? Or do you guys have any suggestions? Thanks a lot!!!\n\nBTW, I'm looking for teammates who have roughly the same or higher score as me (currently 0.0255) and are willing to spend time on this competition (I can spend around 6 hours each day)😀\n\n--------------------\nUpdate: I missed the discussion [here](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/321672#1770555), will try the mentioned method and update the result here.\n\nUpdate2: Seems like the problem still exists", "votes": null }, { "id": "1771242", "postDate": "04/29/2022 03:15:10", "content": "I find a dicussion which may be helpful for you.
\nhttps://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/321672
Thanks a lot!
", "rawMarkdown": "Thanks a lot!", "votes": null }, { "id": "1771824", "postDate": "04/29/2022 15:54:24", "content": "From my understanding the main element that changes the ranker's performance is the negative observations strategy. Each strategy has to be customized for the target group you are looking at - but not to the point of overfitting. The challenge is very complex yet super interesting. You got a nice placement anyway, so you must have gotten something right!
", "rawMarkdown": "From my understanding the main element that changes the ranker's performance is the negative observations strategy. Each strategy has to be customized for the target group you are looking at - but not to the point of overfitting. The challenge is very complex yet super interesting. You got a nice placement anyway, so you must have gotten something right!", "votes": null }, { "id": "1772074", "postDate": "04/29/2022 19:43:17", "content": "thanks for your reply!
", "rawMarkdown": "thanks for your reply!", "votes": null }, { "id": "1774645", "postDate": "05/02/2022 11:02:22", "content": "Our team has currently 0.242 using deep learning models. If you wouldn't mind, we can join your teams and have further communications 😀
", "rawMarkdown": "Our team has currently 0.242 using deep learning models. If you wouldn't mind, we can join your teams and have further communications 😀", "votes": null } ], "comments": [ { "id": 1771242, "author_name": "dengniewei", "author_url": "", "post_date": "04/29/2022 03:15:10", "content": "I find a dicussion which may be helpful for you.
\nhttps://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/321672
Thanks a lot!
", "votes": null, "replies": [] } ] }, { "id": 1771824, "author_name": "lorenzopagliaro01", "author_url": "", "post_date": "04/29/2022 15:54:24", "content": "From my understanding the main element that changes the ranker's performance is the negative observations strategy. Each strategy has to be customized for the target group you are looking at - but not to the point of overfitting. The challenge is very complex yet super interesting. You got a nice placement anyway, so you must have gotten something right!
", "votes": null, "replies": [ { "id": 1772074, "author_name": "weipengzhang", "author_url": "", "post_date": "04/29/2022 19:43:17", "content": "thanks for your reply!
", "votes": null, "replies": [] } ] }, { "id": 1774645, "author_name": "yijunyang123", "author_url": "", "post_date": "05/02/2022 11:02:22", "content": "Our team has currently 0.242 using deep learning models. If you wouldn't mind, we can join your teams and have further communications 😀
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1771078": "My current method contains three steps:\n\n1. retrieve/recall items by some basic rules (buying history, popularity, etc)\n2. label items from step 1 according to the customer's behavior in the valid set\n3. train an lgb ranker to rank the candidates.\n\nThe recall rate of the result in the first step is about 9% (which is really low). However, when I tried to generate more candidates to improve the overall recall rate, the performance of my lgb ranker dramatically decreased, so as the CV score on the valid set😨. Downsampling negative samples couldn't solve the problem.\n\nI'm new to this field, did I miss some essential steps in the process? Or do you guys have any suggestions? Thanks a lot!!!\n\nBTW, I'm looking for teammates who have roughly the same or higher score as me (currently 0.0255) and are willing to spend time on this competition (I can spend around 6 hours each day)😀\n\n--------------------\nUpdate: I missed the discussion [here](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/321672#1770555), will try the mentioned method and update the result here.\n\nUpdate2: Seems like the problem still exists", "1771242": "I find a dicussion which may be helpful for you.\nhttps://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/321672", "1771243": "Thanks a lot!", "1771824": "From my understanding the main element that changes the ranker's performance is the negative observations strategy. Each strategy has to be customized for the target group you are looking at - but not to the point of overfitting. The challenge is very complex yet super interesting. You got a nice placement anyway, so you must have gotten something right!", "1772074": "thanks for your reply!", "1774645": "Our team has currently 0.242 using deep learning models. If you wouldn't mind, we can join your teams and have further communications 😀" }, "source": "meta" }