AENEA: A novel autoencoder-based network embedding algorithm

Abstract

Network embedding aims to represent vertices in the network with low-dimensional dense real number vectors, so that the attained vertices can acquire the ability of representation and inference in vector space. With the expansion of the scale of complex networks, how to make the high-dimensional network represented in low-dimensional vector space through network becomes an important issue. The typical algorithms of current autoencoder-based network embedding methods include DNGR and SDNE. DNGR method trains the Positive Pointwise Mutual Information (PPMI) matrix with the Stacked Denosing Autoencoder (SDAE), which is lacking in depth thereby attaining less satisfactory representation of network. Besides, SDNE used a semi-supervised autoencoder for embedding the adjacency matrix, whose sparsity may generate more cost in the learning process. In order to solve these problems, we propose a novel Autoencoder-based Network Embedding Algorithm (AENEA). AENEA is mainly divided into three steps. First, the random surfing model is used to process the original network to obtain the Probabilistic Co-occurrence (PCO) matrix between the nodes. Secondly, the Probabilistic Co-occurrence (PCO) matrix is processed to generate the corresponding Positive Pointwise Mutual Information (PPMI) matrix. Finally, the PPMI matrix is used to learn the representation of vertices in the network by using a semi-supervised autoencoder. We implemented a series of experiments to test the performance of AENEA, DNGR, SDNE and so on, on the standardized datasets 20-NewsGroup and Wine. The experimental results show that the performance of AENEA is obviously superior to the existing algorithms in clustering, classification and visualization tasks.