I am trying to use DBSMOTE(Density-Based Synthetic Oversampling TEqnique) to on a data set of short text--tweets to be specific. This will be used to train a classifier model in a multiclass classification model. This will be done at feature level augmentation, using TF-IDF as features. I have read though, that to use SMOTE on NLP, the feature vectors must be reduced in dimension. What is an optimal size of a feature vector to use in a SMOTE family algorithm?
Similar Question: How do you apply SMOTE on text classification?
DBSMOTE Code: https://rdrr.io/cran/smotefamily/man/DBSMOTE.html
DBSMOTE Paper: https://link.springer.com/article/10.1007/s10489-011-0287-y