PROJECT TITLE :

Robust Unsupervised Multi-view Feature Learning with Dynamic Graph

ABSTRACT:

By modeling the affinity associations with a graph to lower the dimension, graph-based multi-view feature learning algorithms learn a low-dimensional embedding of the data. The taught low-dimensional representation, on the other hand, is based on a fixed graph that could be erroneous and untrustworthy. Furthermore, the graph creation and projection matrix leaning operations are divided into two discrete steps. We present a robust unsupervised multi-view feature learning method with a dynamic graph to address the issues. The resilient projection matrix is learned at the same time that the dynamic graph structure is built adaptively. We specifically learn a dynamic graph that represents the intrinsic multiple view-specific relations of samples in an adaptable manner. The intrinsic graph structure is preserved and undesirable noises are suppressed using robust projection matrix learning. Furthermore, without any further parameters, the allocated weights are learned automatically for each view. Finally, in order to solve the objective formulation, we devise an efficient alternative optimization procedure. Extensive testing on a variety of multi-view datasets demonstrates the efficacy of our suggested technique.