Generating relation-specific weights for ConvKB using a HyperNetwork architecture

Thanh Le

doi:10.1007/s10489-023-04670-w

Return

Generating relation-specific weights for ConvKB using a HyperNetwork architecture

2023-05-13

PRE

Thanh Le

Duy Nguyen

Bac Le *

DOI：10.1007/s10489-023-04670-w

Original

Original request for help

Save

Abstract

En 中文

Link prediction on knowledge graphs has become increasingly important in various fields, including recommender systems, question answering, and social networks. To address this challenge, ConvKB, a state-of-the-art model in deep learning approaches, has been proposed. However, ConvKB is limited in its ability to exploit general information about relations because it uses the same filters for all relations. Additionally, ConvKB's design, consisting of a convolution layer and a linear layer, may not provide enough parameters to store necessary information during feature learning. To address these limitations, this paper proposes the ConvHyper model, which combines the ConvKB model with a HyperNetwork architecture to create relation-specific weights for convolutional neural network layers. Specifically, the embedding of relations is linearly combined through a HyperNetwork to generate weights for the base neural network. This HyperNetwork architecture helps reduce the complexity of the search space to identify optimal weights and create a relation-specific weight structure for neural network layers. Experimental results show that ConvHyper significantly improves on all metrics in four well-known datasets. ConvHyper improves on H@1 up to 5.5% and achieves better results on other metrics, ranging from 0.8% to 1%. We also compared ConvHyper with other CNN-based models and found that many metrics have better values, particularly from 1.1% to 4.7% on the H@10 metric. Furthermore, we analyzed the influence of hyperparameters and found that the learning rate and negative samples significantly affect the model's results.

Keywords:

Link Prediction

Knowledge Graph

Convolutional Neural Network

HyperNetwork