Robust Unsupervised Multi-View Feature Learning With Dynamic Graph

Abstract

Graph-based multi-view feature learning methods learn a low-dimensional embedding of the data by modeling the affinity correlations with a graph to reduce the dimension. However, the learned low-dimensional representation relies on a fixed graph that is potentially inaccurate and unreliable. Besides, the graph construction and the projection matrix leaning are separated into two independent processes. To tackle the problems, we propose a robust unsupervised multi-view feature learning method with a dynamic graph. The dynamic graph structure is constructed adaptively and the robust projection matrix is learned simultaneously. Specifically, we adaptively learn a dynamic graph which captures the intrinsic multiple view-specific relations of samples. Robust projection matrix learning suppresses the adverse noises and preserves the intrinsic graph structure. Moreover, the assigned weights are learned automatically for each view without any extra parameter. We finally develop an efficient alternative optimization algorithm to solve the objective formulation. The extensive experiments conducted on several multi-view datasets demonstrate the effectiveness of our proposed method. INDEX TERMS Dynamic graph, feature learning, multi-view learning, robust projection matrix.

Matching of 3D Model and Semantic Description via Multi-Modal Auxiliary Classifier Generative Adversarial Network With Autoencoder

“…For the latent correlated representation of sharing network output, to preserve the relationships between data after feature projection [37], [38], this paper defines a structure-preserving loss which is based on the triple loss. The equation of the triple loss is as follows:…”