Dealing with Imbalanced Datasets in ML Classification Problems | DataHour by Damini Dasgupta

[@alexilaiho6441]

The talk fails to cover how to deal with imabalnced datasets using SMOTE, and also using Focal Loss for Neural Nets

[@Analyticsvidhya]

DataHour Resources 🔗 bit.ly/3xUaue4

[@shreyashimukhopadhyay6354]

Great tutorial on imbalanced data 👍Analytics Vidya Can you please share the notebook.

[@Analyticsvidhya]

Dear Shreyashi, here's the download link: drive.google.com/drive/folders/1KV-CNZmuy8sqYDc1Jri-hCvYCeHSbYxs?usp=sharing

[@shreyashimukhopadhyay6354]

@@Analyticsvidhya Thank you so much! 👍

[@younesgasmi8518]

thank you so much miss. my question is can I use the undersampling technique before splitting the dataset into training and testing sets because there is not any data leakage when we use this method

[@Analyticsvidhya]

You're absolutely right that undersampling can be done before splitting the dataset in imbalanced classification problems. In fact, it's generally considered the preferred approach to avoid data leakage! Here's why: ➡️ Data leakage: If you undersample after splitting, some minority class instances might accidentally leak into the test set, making your model's performance seem better than it truly is on unseen data. ➡️ Preserving real-world distribution: Undersampling before the split ensures the training set reflects the actual imbalance in your real-world data, leading to a model that generalizes better. However, keep in mind that undersampling also has drawbacks like losing potentially valuable minority class data. It's always a good idea to compare different techniques like oversampling or SMOTE before making a final decision.

[@gecarter53]

Great seminar. Is the code publicly available?

[@Analyticsvidhya]

Dear Learner, Refer to our DataHack Platform for DataHour Material and Speaker Coordinates: datahack.analyticsvidhya.com/contest/all/

[@chukwumanwakpa3330]

Thank you so much analytics Vidhya. I must commend you all for my improvement in deploying ML algorithms in solving problems. Just an observation please, I think it would be better if we can get the slide for all the data sets used in all lectures. thank you. Much love from Nigeria.

[@Analyticsvidhya]

Dear Learner, Refer to our DataHack Platform for DataHour Material and Speaker Coordinates: datahack.analyticsvidhya.com/contest/all/

[@therevolution8611]

Can I use oversampling , if I have multi label text for classification purpose??

[@Analyticsvidhya]

Dear Learner, Refer to our DataHack Platform for DataHour Material and Speaker Coordinates: datahack.analyticsvidhya.com/contest/all/