Multi-label Stream Classification Using Extended Binary Relevance Model

In this paper the issue of multi-label data stream classification was addressed. To deal with the posed problem, we introduced a recognition system that is build upon a two level architecture. The first level is a Binary Relevance multi-label classifier, and the second is a correction procedure that employs competence and cross-competence measures to adjust the output of the Binary Relevance classifier. The measures are calculated in a lazy manner using a local, fuzzy confusion matrix which is a generalized version of commonly known confusion matrix. The matrix follows the changes in the data stream by using a sliding window approach. The sliding window is implemented as a simple FIFO queue. During the experimental study the algorithm was compared against 2 state-of-the-art approaches using 12 benchmark datasets. The classification quality of the investigated model was performed using 8 different evaluation measures. The conducted the experiments revealed that the proposed method outperforms the reference methods in terms of the Hamming loss and micro-averaged False Discovery Rate. On the other hand, the experimental outcome suggests that, in general, the corrected base classifiers of the BR ensemble are biased towards the majority class.