Dynamic network embedded link prediction method based on variational auto-encoder

Abstract

The invention discloses a dynamic network embedded link prediction method based on a variational auto-encoder, which comprises the following steps of S1, obtaining academic network or social network data streams for a period of time, and performing data preprocessing according to a certain time fine granularity; S2, establishing an encoding and decoding framework by using a variational auto-encoder, and encoding the network data of each time slice to obtain a low-dimensional vector; and S3, applying a self-attention constraint to the low-dimensional vector output by each node in the network byutilizing the neighborhood between the nodes in the current network. According to the method, the core mechanism of dynamic network evolution is discovered by mining the potential relationship between the original network and the implicit vector, so that the information of the network in space and time is effectively utilized, the defects of the traditional static method are overcome, and the intermediate implicit vector generated by the auto-encoder is constrained; spatial information and time information of the network are effectively utilized.

Description

technical field [0001] The invention relates to the technical field of deep learning model methods, in particular to a dynamic network embedding link prediction method based on a variational autoencoder. Background technique [0002] Complex systems can be represented as networks, such as protein networks, social networks, communication networks, and co-author networks. The subtasks of network analysis usually include node classification, community discovery, link prediction, recommendation systems, etc. Network representation learning or network embedding is to reduce the complexity of the original network and improve the research efficiency of downstream tasks by mapping the high-dimensional original network into a low-dimensional space, while the representation learning of the dynamic network can use the historical information of the network to predict the next The Evolution of Moment Networks. [0003] Representation learning for complex networks can map a high-dimensio...

AI Technical Summary

Problems solved by technology

[0004] However, in the real world, the network often evolves dynamically, and the network that only focuses on a certain static moment tends to ignore the evolution of the network, making the application scenarios narrow

Based on this, some representation learning methods for dynamic networks pay more attention to the evolution of the network, making the classification and link prediction of models in dynamic scenarios better than traditional static methods. These dynamic methods make up for the shortcomings of static methods in describing network changes, but The essential mechanism of network evolution is still not explained, nor is the essential connection between the original network and the hidden vector found

Images