Global-local neighborhood based network representation for citation recommendation

Many researchers study citation recommendation approaches using network representation recently. It learns low-dimensional vector representation of nodes in a citation network, generates a recommendation list using similarity scores within the obtained node representations. Most of the existing approaches learn network representation by preserving structure information ofthe citation network. However, nodes in a citation network often associated with content information, recent approaches tend to learn each node’s structure and content representations separately, and apply a simple and empirical combination strategy to produce the final node representation vectors which are suboptimal. To solve the above problems, we propose a Global-Local Neighborhoods based Network Representation model, named GLNNR, to integrate network structure and non- structural information, obtaining better node representation in the network. For global neighborhoods, we first generate parallel sequences containing node identity sequences and the corresponding content sequences, then we propose an attention- based sequence to sequence component to obtain node embeddings (the learned hidden representations of encoder) under global neighborhoods. For local neighborhoods, inspired by matrix factorization, a component can be designed to fuse structural and non-structural information in the lower-order neighborhood, here we first build a co-occurrence matrix (adjacent matrix) of the network, and then use multilayer perceptron to learn node representations under the lower-order neighborhood. The final node representations are global-local neighborhood based node embeddings. Empirical experiments prove the effectiveness of the GLNNR on a real-world information citation networks, i.e. CiteSeer and DBLP. Besides, we conduct experiments on a web page citation network, i.e. Wikipedia, to prove extensibility and portability of the proposed model.