A Method for Constructing a Cross-Media Heterogeneous Information Network

The underlying heterogeneity leads to a serious semantic gap between cross-media data. To solve this problem, research on cross-media relevance calculation and cross-media retrieval has flourished. However, current cross-media retrieval technology still has problems such as poor retrieval practicality, messy data, and high looseness. To address these problems, this paper proposes a cross-media heterogeneous information network construction method to define and construct a heterogeneous information network of cross-media data. First, the network form of cross-media heterogeneous information is clarified, and the nodes and edges of the network are constructed using existing cross-media association learning methods and technologies. Then, in order to streamline the network structure, a threshold selection algorithm based on genetic algorithm to solve the maximum average precision mean is proposed to prune the existing network, thereby further improving the efficiency and applicability of the network. Experiments are conducted on XMediaNet, the world's largest cross-media dataset, to verify the effectiveness of the proposed method.