Flexible data representation with feature convolution for semi-supervised learning

Abstract

Data representation plays a crucial role in semi-supervised learning. This paper proposes a framework for semi-supervised data representation. It introduces a flexible nonlinear embedding model that integrates graph-based data convolutions. The proposed approach exploits structured data in order to estimate a nonlinear data representation as well as a linear transformation, enabling an inductive semi-supervised model. The introduced approach exploits data graphs at two different levels. First, it integrates manifold regularization that is encoded by the graph itself. Second, it optimizes a flexible linear transformation that maps the convolved data samples to their nonlinear representations. These convolved data are generated by the joint use of the graph and data. The proposed semi-supervised model overcomes some challenges related to some samples distributions in the original spaces. The proposed Graph Convolution based Semi-supervised Embedding (GCSE) provides flexible models which can improve both the data representation and the final performance of the learning model. Experiments are run on six image datasets for comparing the proposed approach with several state-of-art semi-supervised methods. These results show the effectiveness of the proposed framework.