A Probability-Based Close Domain Metric in Lifelong Learning for Multi-label Classification

Lifelong machine learning has recently become a hot topic attracting the researchers all over the world by its effectiveness in dealing with current problem by exploiting the past knowledge. The combination of topic modeling on previous domain knowledge (such as topic modeling with Automatically generated Must-links and Cannot-links, which exploits must-link and cannotlink of two terms), and lifelong topic modeling (which employs the modeling of previous tasks) is widely used to produce better topics. This paper proposes a close domain metric based on probability to choose valuable knowledge learnt from the past to produce more associated topics on the current domain. This knowledge is, then, used to enrich features for multi-label classifier. Several experiments performed on review dataset of hotel show that the proposed approach leads to an improvement in performance over the baseline.