A method for embedding temporal knowledge graphs using graph-core attention networks

Abstract

The present application relates to a graph kernel attention network temporal knowledge graph embedding method, which takes dishonest entities as nodes and dishonest relations as edges, and calculates weights based on dishonest relations and timestamps to construct a directed graph; uses a graph attention network to extract feature representations of nodes in the directed graph; applies a twin neural network to perform correlation analysis on the feature representations of nodes and extracts similarity relationship features; enhances the attention mechanism to calculate global dishonest entity relationship weights based on dishonest relations and similarity relationship features; maps the feature representation of nodes to a high-dimensional space by calculating a kernel function through a domain graph kernel; calculates the feature entropy value of the updated weight through linear structural entropy; obtains a high-dimensional embedding vector of the node by combining the domain graph kernel processing result and the feature entropy value; uses the t‑SNE method to reduce the dimension of the high-dimensional embedding vector of the node; and fuses the reduced high-dimensional embedding vectors of dishonest entities corresponding to multiple timestamps and dishonest relations to obtain an embedding vector representation.

Images