What are some of the techniques fellow Kagglers have used in the past to tackle class imbalance in the multi-label classification scenarioes?
", "messages": [ { "id": 1252298, "postDate": "2021-03-25T15:09:23.717Z", "content": "What are some of the techniques fellow Kagglers have used in the past to tackle class imbalance in the multi-label classification scenarioes?
", "rawMarkdown": "What are some of the techniques fellow Kagglers have used in the past to tackle class imbalance in the multi-label classification scenarioes?", "votes": 3 }, { "id": 1255395, "postDate": "2021-03-28T18:44:56.437Z", "content": "Hi there!
\nYou can check the previous HPA competition discussion board for some pretty awesome insights.
\nI believe many used Focal Loss w/ great success when attempting to mitigate the class imbalance problems associated with the multi-label classification of slide-level label prediction. In fact, here is a paper titled Focal Loss Improves the Model Performance on Multi-Label Image Classifications with Imbalanced Data that supports this hypothesis.
\nI will caveat all of this by saying that I have not had any luck with achieving higher scores by using Focal Loss in this competition.
\nOther techniques might include (but are not limited to):
\nHey @dschettler8845 thank you for the reply.
\nI have tried with focal loss as well but to no luck. Currently I am creating cell level dataset to mitigate issues of class imbalance using proper data. Let's see how it plays out.
\nOversampling/undersampling is good but got it's own caveats. Trying to avoid it as much as possible.
", "rawMarkdown": "Hey @dschettler8845 thank you for the reply. \n\nI have tried with focal loss as well but to no luck. Currently I am creating cell level dataset to mitigate issues of class imbalance using proper data. Let's see how it plays out. \n\nOversampling/undersampling is good but got it's own caveats. Trying to avoid it as much as possible. \n", "votes": 1 } ] } ], "comments": [ { "id": 1255395, "author_name": "Darien Schettler", "author_url": "", "post_date": "2021-03-28T18:44:56.437000", "content": "Hi there!
\nYou can check the previous HPA competition discussion board for some pretty awesome insights.
\nI believe many used Focal Loss w/ great success when attempting to mitigate the class imbalance problems associated with the multi-label classification of slide-level label prediction. In fact, here is a paper titled Focal Loss Improves the Model Performance on Multi-Label Image Classifications with Imbalanced Data that supports this hypothesis.
\nI will caveat all of this by saying that I have not had any luck with achieving higher scores by using Focal Loss in this competition.
\nOther techniques might include (but are not limited to):
\nHey @dschettler8845 thank you for the reply.
\nI have tried with focal loss as well but to no luck. Currently I am creating cell level dataset to mitigate issues of class imbalance using proper data. Let's see how it plays out.
\nOversampling/undersampling is good but got it's own caveats. Trying to avoid it as much as possible.
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "1252298": "What are some of the techniques fellow Kagglers have used in the past to tackle class imbalance in the multi-label classification scenarioes?", "1255395": "Hi there!\n\n---\n\nYou can check the previous [**HPA competition discussion board**](https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion) for some pretty awesome insights.\n\nI believe many used **Focal Loss** w/ great success when attempting to mitigate the class imbalance problems associated with the multi-label classification of slide-level label prediction. In fact, here is a paper titled [**Focal Loss Improves the Model Performance on Multi-Label Image Classifications with Imbalanced Data**](https://dl.acm.org/doi/10.1145/3411016.3411020) that supports this hypothesis. \n\nI will caveat all of this by saying that I have not had any luck with achieving higher scores by using Focal Loss in this competition.\n\n---\n\nOther techniques might include (but are not limited to):\n- [**Oversampling or Undersampling**](https://machinelearningmastery.com/random-oversampling-and-undersampling-for-imbalanced-classification/)\n- Using External Data To Bolster the Rare Classes --> [**Useful Dataset By Alexander Riedel**](https://www.kaggle.com/alexanderriedel/hpa-public-768-excl-0-16) ( @alexanderriedel )" } }