{ "id": 305179, "title": "Doubt about rare Individual Ids ", "url": "/competitions/happy-whale-and-dolphin/discussion/305179", "author_name": "", "post_date": "2022-02-04T05:43:25.113628500Z", "votes": 7, "comment_count": 2, "views": 0, "content": "

Greetings everyone. I want to ask how are you all using individual ids for training which are present just once or twice in the whole dataset ? Should we remove those rare ids ?

", "messages": [ { "id": "1675227", "postDate": "02/04/2022 05:43:25", "content": "

Greetings everyone. I want to ask how are you all using individual ids for training which are present just once or twice in the whole dataset ? Should we remove those rare ids ?

", "rawMarkdown": "Greetings everyone. I want to ask how are you all using individual ids for training which are present just once or twice in the whole dataset ? Should we remove those rare ids ?", "votes": null }, { "id": "1675452", "postDate": "02/04/2022 09:22:52", "content": "

I would not recommend removing rare individuals. You can make a profit from even 1 sample per individual. For example, you could try to use something like contrastive loss (check out this article).
\nHere is a good starter notebook by @debarshichanda - [Pytorch] HappyWhale Siamese Starter which implements siamese network.

", "rawMarkdown": "I would not recommend removing rare individuals. You can make a profit from even 1 sample per individual. For example, you could try to use something like contrastive loss ([check out this article](https://medium.com/@maksym.bekuzarov/losses-explained-contrastive-loss-f8f57fe32246)). \nHere is a good starter notebook by @debarshichanda - [[Pytorch] HappyWhale Siamese Starter](https://www.kaggle.com/debarshichanda/pytorch-happywhale-siamese-starter) which implements siamese network.", "votes": null }, { "id": "1678646", "postDate": "02/06/2022 17:25:00", "content": "

The previous Kaggle competition had a large/similar class imbalance. You should also checkout the solutions from there.

", "rawMarkdown": "The previous Kaggle competition had a large/similar class imbalance. You should also checkout the solutions from there.", "votes": null } ], "comments": [ { "id": 1675452, "author_name": "meowmeowmeowmeowmeow", "author_url": "", "post_date": "02/04/2022 09:22:52", "content": "

I would not recommend removing rare individuals. You can make a profit from even 1 sample per individual. For example, you could try to use something like contrastive loss (check out this article).
\nHere is a good starter notebook by @debarshichanda - [Pytorch] HappyWhale Siamese Starter which implements siamese network.

", "votes": null, "replies": [] }, { "id": 1678646, "author_name": "init27", "author_url": "", "post_date": "02/06/2022 17:25:00", "content": "

The previous Kaggle competition had a large/similar class imbalance. You should also checkout the solutions from there.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1675227": "Greetings everyone. I want to ask how are you all using individual ids for training which are present just once or twice in the whole dataset ? Should we remove those rare ids ?", "1675452": "I would not recommend removing rare individuals. You can make a profit from even 1 sample per individual. For example, you could try to use something like contrastive loss ([check out this article](https://medium.com/@maksym.bekuzarov/losses-explained-contrastive-loss-f8f57fe32246)). \nHere is a good starter notebook by @debarshichanda - [[Pytorch] HappyWhale Siamese Starter](https://www.kaggle.com/debarshichanda/pytorch-happywhale-siamese-starter) which implements siamese network.", "1678646": "The previous Kaggle competition had a large/similar class imbalance. You should also checkout the solutions from there." }, "source": "meta" }