Multi-label classification is a generalization of a multi-class classification problem where one entity can belong to more than one class from the class set. Recent works have proposed multiple methods of solving this problem that involves both statistical and deep learning methods. While methods exist for using deep learning models for this problem, most of them require the model to have a high dimension output vector and the property of inter-dependency of classes has not been explored. An ensemble of statistical models called the chain classifiers can be used to address these issues. This study explores methods of using neural network classifiers in the classifier chain model and tries to address some problems with such architecture while compare their performance on different types of data using different metrics with each other and with other well performing multi-label classification methods.
This model requires feeding the output node of the NN for lable one to the input vector of the NN for the next lable,
thus considering the dependency.
A simple chain architecture doesn't consider the direction of considering dependency. The lable Kn will have the output from lables
K1 to Kn-1 in consideration whereas the first lable K1 will have none. Therefore, bidirectional chains create chains in both the directions
and ensemble thier results to get a more accurate prediction.
Training a neural nets based CC can be taxing. However, using the fact that the middle layer weights in subsequent classifiers of a chain has almost similar weights, we can transfer the weights from the first classifer to the following ones in order to reduce the number of total nodes requiring backpropagation.
