Multimodality Fusion for Node Classification in D2D Communications

With the rapid development of 5G technology, D2D communications will be widely used in content distribution scenarios. An important prerequisite for realizing data transmission in D2D scenarios is to correctly classify each node in the network. However, most of existing node classification methods need huge computational overhead in large-scale networks. Inspired by network structure and attribute information, a multimodality fusion-based node classification scheme is proposed to obtain the tradeoff between classification accuracy and costs. First, in order to summarize label dependences to perform node classification, structure and information features are constructed. Structure features are obtained by embedding network data into a low-dimensional representation space, and information features are extracted from attribute information. Second, a multimodality fusion model is applied to integrate knowledge learned from different modalities, which can further improve the overall performance. Besides, the multilayer neural network based on softmax regression is used for multi-label classification. Finally, extensive experiments on two public datasets are conducted to verify the validity of the proposed scheme. Specifically, experimental results demonstrate the superior performance of our method over existing solutions in terms of accuracy and overhead.