Lifelong Spectral Clustering.

摘要: In the past decades, spectral clustering (SC) has become one of most effective algorithms. However, previous studies focus on tasks with a fixed task set, which cannot incorporate new without accessing to previously learned tasks. this paper, we aim explore problem in lifelong machine learning framework, i.e., Lifelong Spectral Clustering (L2SC). Its goal is efficiently learn model for by selectively transferring accumulated experience from knowledge library. Specifically, library L2SC contains two components: 1) orthogonal basis library: capturing latent cluster centers among clusters each pair tasks; 2) feature embedding manifold information shared multiple related As arrives, firstly transfers both and obtain encoding matrix, further redefines base over time maximize performance across all Meanwhile, general online update formulation derived alternatively Finally, empirical experiments several real-world benchmark datasets demonstrate that our can effectively improve when comparing other state-of-the-art