Binary relevance for multi-label learning: an overview