Design of multi-view graph embedding using multiple kernel learning

Abstract

The graph embedding is the process of representing the graph in a vector space using properties of the graphs and this technique has now being widely used for analyzing the graph data using machine learning algorithms. The existing graph embeddings rely mostly on a single property of graphs for data representation which is found to be inappropriate to capture all the characteristics of the data. Hence we designed graph embedding using multi-view approach, where each view is an embedding of the graph using a graph property. The input space of multi-view learning is then taken as the direct sum of the subspaces in which the graph embedding lie. We did analysis on real world data by incorporating the proposed model on support vector machines (SVM). The reproducing kernel used in SVM is represented as the linear combination of the kernels defined on the individual embeddings. The optimization technique used in simple multiple kernel learning (simpleMKL) is used to find the parameters of the optimal kernel. To analyze the individual representation capability of the embeddings, an R-convolution graph kernel is designed over each of the views. In our experimental analysis, the multi-view graph embedding showed a superior performance in comparison with that of the state-of-the-art graph embeddings as well as graph kernels.