Abstract:
As a default graph dataset, the node graph is often used in network analysis. By employing this method, the vector representation of nodes can be obtained, but the edges cannot be directly represented as vectors. In this paper, a graph embedding algorithm named Line2Vec is proposed to combine biased sampling with unbiased sampling. And on this basis, a link prediction framework named Line2Vec-L is suggested. Firstly, Line2Vec redefines the sampling probability of nodes in sampling domain based on the above comprehensive walk strategy. And by combining Word2Vec model, the enriching edge-embedding vector representation of node graph is obtained to avoid impairing the information. Then. the vector representation for non-existing edges and unknown edges can be learned with incidence matrix. All the obtained vectors of edges are applied to link prediction. Experiments show that Line2Vec is effective on edge-vector representation and the performance of AUC with Line2Vec-L is better. In other words, it means the quality of edge-vector representation generated by Line2Vec is higher, which helps to improve the performance of link prediction.