Abstract

Most existing network representation algorithms learn the network representations based on network structure, however, they neglect the rich external information associated with nodes (i.e. text contents, communities and label information). Meanwhile, the learnt representations usually lack the discriminative ability for the tasks of node classification and linking prediction. We consequently overcame the above challenges by presenting a novel semi-supervised algorithm, text-associated max-margin DeepWalk algorithm (TMDW). TMDW incorporates text contents and network structures into the network representation learning based on the inductive matrix completion algorithm, and then we use node’s category to optimize the learnt network representations based on the mar-margin principle and biased gradient. For integrating the above tasks, we propose a novel and efficient framework of network representation learning, this framework is easy to extend and generate discriminative representations. We then evaluate our model using the multi-class classification tasks. The experimental results demonstrate that TMDW outperforms other baseline methods on three real-world datasets. The visualization task of TMDW shows that our model is more discriminative than the other unsupervised approaches.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call