Structural iterative lexicographic autoencoded node representation

Graph representation learning approaches are effective to automatically extract relevant hidden features from graphs. Previous related work in graph representation learning can be divided into connectivity and structural-based. Connectivity-based representation learning methods work on the assumption that neighboring nodes should have similar representations. While structural node representation learning assumes that nodes with the same structure should have identical representations; structural representation learning is suitable for node classification and regression tasks. Possible drawbacks of current structural node representation learning approaches are prohibitive execution time complexity and the inability to entirely preserve structural information. In this work, we propose SILA, a Structural Iterative Lexicographic Autoencoded approach for node representation learning. This new iterative approach presents a small number of iterations, and compared with the method presented in the literature, shows better performance in preserving structural information for both classification and regression tasks.