Lifelong Hierarchical Topic Modeling via Non-negative Matrix Factorization

Hierarchical topic modeling has been widely used in mining the latent topic hierarchy of documents. However, most of such models are limited to a one-shot scenario since they do not use the identified topic information to guide the subsequent mining of topics. By storing and exploiting the previous knowledge, we propose a lifelong hierarchical topic model based on Non-negative Matrix Factorization (NMF) for boosting the topic quality over a text stream. In particular, we construct a knowledge graph by the accumulated topic hierarchy information and use the knowledge graph to guide the training of our model on future documents. Moreover, the structure information in the knowledge graph is completed by supervised learning. Experiments on real-world corpora validate the effectiveness of our approach on lifelong learning paradigms.