关键词: 多标签分类, 嵌入空间学习, 变分自动编码器, Transformer, 标签相关性

Abstract: With the rise of the Internet, all kinds of data are growing rapidly, and how to utilize these sample data efficiently has become an important issue in the field of data mining. The multi-label classification task, as an important task in the field of machine learning and data mining, aims to label samples with multiple label categories. Most of the current methods only learn embedding representations for feature branches, do not take into account the semantic relevance between features and labels, and lack effective constraints on the feature embedding space, which leads to insufficient relevance of the learnt feature embeddings. Meanwhile, in terms of label relevance learning, most of the existing methods mainly focus on low-order label relevance, and thus the problem of insufficient learning of high-order relevance between multiple labels becomes more prominent when facing complex actual labeling scenarios. Therefore, in order to solve the above problems, this paper proposes a multi-label classification method based on label-aware variational self-encoder from embedding representation learning and label relevance learning. Specifically, for embedding representation learning, this paper proposes to use feature and label dual-stream variational self-encoders to simultaneously learn and align the embedding space of features and labels, and add label guidance to the feature embedding space to enhance feature embedding. Meanwhile, a label semantic-based cross-attention mechanism is used to add specific label information to the feature embedding, and finally discriminative feature embeddings after label sensing are obtained. For label relevance learning, the multi-layer self-attention mechanism in the shared decoder is used to fully fuse the similarity information of multiple labels, and through the co-occurring interactions between different labels, the label higher-order relevance representations are learnt and used for cross-aware feature embedding. Experimental results obtained on datasets from four different domains show that the proposed method can effectively enhance feature and label embedding and fully capture the higher-order correlation information between labels for multi-label classification tasks, and the significant superiority of the proposed method in performance is verified through a comparative analysis with state-of-the-art algorithms in terms of multiple evaluation metrics.

Key words: multi-label classification, embedded space learning, variational autoencoder, Transformer, label correlation