Online learning for low-rank representation and its application in subspace clustering

Subspace clustering is an important problem in machine learning and computer vision research. The low-rank representation (LRR) model is an extension of the famous robust principle component analysis and provides a state-of-the-art solution to the subspace clustering problem. We propose in this paper the first online learning algorithm for solving the LRR model. The traditional batch algorithms for LRR suffers from high time complexity and space complexity, and their scalability is largely limited. By comparison, our proposed online learning algorithm processes one data instance at a time, thus it can significantly reduce the time and space costs and can apply to large-scale data. The experiments on simulated data and real-world data all demonstrate the effectiveness and efficiency of our online learning algorithm. 1553-9105/Copyright © 2014 Binary Information Press