Fusing heterogeneous information for multi-modal attributed network embedding

In the real world, networks with many types of nodes and edges are complex, forming a heterogeneous network. For instance, film networks contain different node types, such as directors, films and actors, as well as different types of edge and multimodal attributes. Most existing attribution network embedding algorithms cannot flexibly capture the impact of multimodal attributes on the topology. Premature fusion of multimodal features encodes different attribute information into the representation embedding, while the later fusion strategy ignores the interaction between different modes, both of which affect the modeling of graph embedding.To solve this problem, we propose a multimodal attribute network representation learning algorithm based on heterogeneity information fusion, named FHIANE. It extracts features from multimodal information sources through deep heterogeneous convolutional networks and projects them into a consistent semantic space while maintaining structural information. In addition, we design a modality fusion network based on an extended attention mechanism that takes full advantage of the consistency and complementarity of multimodal information. We evaluate the performance of the FHIANE algorithm on several real datasets through challenging tasks such as link prediction and node classification. The experimental results show that FHIANE outperforms other baselines.