Topology Learning Embedding: A Fast and Incremental Method for Manifold Learning

摘要

In this paper, we propose a novel manifold learning method named topology learning embedding (TLE). The key issue of manifold learning is studying data's structure. Instead of blindly calculating the relations between each pair of available data, TLE learns data's internal structure model in a smarter way: it constructs a topology preserving network rapidly and incrementally through online input data; then with the Isomap-based embedding strategy, it achieves out-of-sample data embedding efficiently. Experiments on synthetic data and real-world handwritten digit data demonstrate that TLE is a promising method for dimensionality reduction.