An entity-weights-based convolutional neural network for large-sale complex knowledge embedding

Highlights：

• In terms of complex relationships and multi-relational circle structures, we demonstrate that they are ubiquitous and inevitable.

• The measures selected for baseline methods are not optimal by comparison experiments, and we develop two models based on matrix factorization, called sMFE and tMFE to improve embedding results.

• We first use the complete incidence matrix to construct ComInE model, which contain the most comprehensive topological properties of graphs.

• Use entity weights calculated by PageRank to extend TransE and apply them in Deep Learning, which reduce the cost of adding extra information while ensuring the performance.

• A convolutional neural network with three hidden layers based on node weights, CNNe, is proposed for knowledge embedding.