I'm trying to solve a multilabel classification problem with n-binary LSTM classifiers. I have 17 classes in total, where multiple classes may be true for each example (e.g. news articles with multiple correct categories).
The classification is very successful with some classes, which have about 1/10th true positive rate, but for classes with a rate of 1/100th, 1/1000th, the classifier doesn't learn much -- everything ends up becoming true negatives.
Since the classifiers for the different models are independent, I thought about manufacturing tailored datasets with higher rates for each classifier. Is this a bad idea?
For the record; I'm using Word2Vec vectors as features for the input, and the network itself is a shallow with 2-layers and 333 hidden nodes. Training is done with backprop and SGD, with a minibatch size of 100. I can provide more info, but I'm not sure how relevant it is.
Any suggestions are welcome.
