Knowledge graph geometric model expression ability enhancement method based on trap function

Abstract

The invention discloses a knowledge graph geometric model expression ability enhancement method based on a trap function, and belongs to the field of knowledge graph expression learning algorithms. The method specifically comprises the following steps: firstly, giving a geometric knowledge graph representation learning model S and a relation mode P, establishing an n-element linear equation set, solving the linear equation set to obtain a relation constraint Rc and an optimal solution O under the relation mode P, then, constructing a trap function near the optimal solution, and constructing related loss and zero constraint loss of the relation constraint Rc based on the trap function, so as to obtain a geometric knowledge graph representation learning model; adding into a loss function of a given geometric knowledge graph representation learning model S to serve as a final loss function of the model; and finally, optimizing the model by adopting a gradient descent algorithm to obtain the final representation of the model S, thereby remarkably improving the capability of the model in expressing the relation mode and linking the prediction task. According to the method, the expression effect of the model is improved, and the high efficiency of model training is ensured.

Images