Abstract
Most of existing community detection algorithms group nodes with more connections into the same community, and they are more concerned with links within the community. However, the weak ties between different communities are also important, because they can reflect the relationships between different communities, including helpful, friendly or negative, and adverse. Few studies focus on weak ties, although they are important. In this paper, we propose a novel sign prediction model based on the nodes features in the network, including the Jaccard similarity and the ratio of the negative degrees of all nodes, and the autoencoder technology that self-defines its loss function with the features of the communities. The proposed model maps the original network to a low-dimensional space so that the weak ties can be represented by low-dimensional vectors. We conduct experiments on the Epinions and Slashdot datasets and find that the proposed model outperforms the challenging state-of-the-art graph embedding methods in the sign prediction of weak ties in terms of accuracy and F1 score measurement.