Reasearch on feature mapping based on labels information in multi-label text classification

Feature representation plays an important role in text classification. Feature mapping based on labels information is an algorithm suitable for Binary Relevance. Compared with the conventional text representation, it makes the dimension of the text under control by means of word embedding. More importantly, it takes full advantage of the general characteristics of the label on text representation. In addition, it does not require feature extraction and dimension reduction, which ensure the integrity of text information. First, it applies K-means clustering to positive and negative samples of each label so as to extract the general characteristics. And then, each word embedding in a text is assigned a different weight according to the distance with all clustering centers. In this way, each classifier maps a same text to a different matrix relevant to the corresponding label clustering center. The experiment shows that proposed algorithm provides better performance than conventional feature representation on all evaluation standards.