I have a multi-class problem. I want to solve this problem using the LCPN (Local Classifier per Parent Node) method of heirachical classification. Therefore, I have manually divided the original dataset into a hierarchy of 2 levels. Level 1 has, four parent classes & each of the parent classes have child classes (level 2). Does anyone know if the below mentioned approach is correct to solve this problem?
for each of the 4 sub classifiers under the parent classifier:
4. take the resampled training set (the dataset obtained after resampling, at step 2. )
5. filter the instances that belong only to the particular sub classifier. Because the resampled dataset has instances that belong to all the 4 sub classifiers
6. create a dataset using those instances. Including features & labels
7. split the created dataset, into sub-training(70%) and sub-test set(30%)
8. oversample the sub-training set (because the classes are imbalanced)
9. train the model with the oversampled sub-training set at step 8., & test it using the sub-test set
Dear Sanushi Salgado ,
For your query, I suggest you to follow these links: https://towardsdatascience.com/hierarchical-classification-with-local-classifiers-down-the-rabbit-hole-21cdf3bd2382
https://towardsdatascience.com/https-medium-com-noa-weiss-the-hitchhikers-guide-to-hierarchical-classification-f8428ea1e076
https://datascience.stackexchange.com/questions/46329/hierarchical-classification-with-multi-class-predictor-for-every-parent-node
I didn't understand why you are complicated the training process. From the beginning you can divide your database into smaller ones such that each sub-database will hold the classes of the given level. After that, your training steps can be easily applied for each sub database separately.
Thank you @Suhaila for suggesting that approach. According to what you have said, I hope that I should perform the following steps.
1. filter intances that belong to the classes, classified by each sub classifier & put them in seperate datasets
2. train the parent classifier on the entire dataset
3. then train each of the sub classifiers on the corresponding datasets that contain the filtered instances
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training