Inductive Learning of Dynamic Graph Embeddings

In this paper we research the problem of learning graph embeddings with respect to its applications in dynamic networks and ability to generalize to unseen nodes. Firstly, we overview the state-of-the-art methods and techniques of retrieving graph embeddings and learning algorithms for both transductive and inductive approaches. Secondly, we propose an improved model GSM based on GraphSAGE algorithm and set up the experiments on datasets CORA, Reddit, and HSEcite —— our own data collected from Scopus citation database across authors with affiliation to NRU HSE in 2011-2017. The results show that our three-layer model with attention-based aggregation function, added normalization layers, regularization (dropout) outperforms GraphSAGE models with mean, LSTM and pool aggregation functions taken as baselines.