A metrics-based meta-learning model with meta-pretraining for industrial knowledge graph construction

Abstract

Isolated data silos and domain-specific knowledge pose challenges for knowledge graph construction in the manufacturing industry, where heterogeneous storage leads to distributed databases with complex schemas. In this article, a resource-based industrial knowledge graph is developed using a few-shot classification algorithm to save on labor and other related costs in industrial knowledge graph construction, and an attribute-based fusion strategy for data fusion and alignment is designed. We also propose a novel metrics-based meta-learning model with meta-pretraining (MMM) to address the few-shot text classification problem. Experiment results show that MMM achieves 87.13% accuracy on the 5-shot text classification benchmark Amazon Review Sentiment Classification (ARSC), outperforming other baselines, such as Induction Networks (85.63%) and Distributional Signatures (81.16%). The MMM achieves a 34.6% accuracy improvement compared with Distributional Signatures (84.34% vs. 62.66%) on 1-shot problems of ARSC, hence highlighting the applicability of our model in low-resource conditions. Based on the proposed methods, we further develop an industrial knowledge graph platform with industrial applications, such as value chain analysis and collaboration, to improve knowledge reuse and service innovation.