Semi-supervised Learning Algorithm for Binary Relevance Multi-label Classification

The presented paper describes our model for the WISE 2014 challenge multi-label classification task. The goal of the challenge was to implement a multi-label text classification model which maximizes the mean F-1 score on a private test data. The described method involves a binary relevance scheme with linear classifiers trained using stochastic gradient descent. A novel method for determining the values of classifiers' meta-parameters was developed. In addition, our solution employs the semi-supervised learning which significantly improves the evaluation score. The presented solution won the third place in the challenge. The results are discussed and the supervised and semi-supervised approaches are compared.