A Robot Control Knowledge Recommendation Model PKGAT Based on Multimodal Knowledge Graph

In the field of robotic control, the interdisciplinary and complex nature of knowledge leads to issues such as knowledge fragmentation and a steep learning curve, posing significant challenges to mastering domain expertise. Recommendation systems, as typical information-filtering tools, are often employed to facilitate knowledge retrieval. Nevertheless, they inherently suffer from cold-start and data sparsity problems, which compromise recommendation accuracy. To overcome these limitations, this study first combines the recommendation system with the knowledge graph and improves upon the traditional BERT-BiLSTM-CRF entity recognition model by incorporating an attention mechanism to enhance entity extraction performance. Verified on the public dataset, the accuracy rate, recall rate and F1 value of the improved model were 91.6%, 94.5% and 93.1%, respectively, which increased by 1.4%, 2.9% and 2.3%, respectively, compared with the BERT-BiLSTM-CRF model. Subsequently, by using the YOLOv5 object detection network to extract image features from the self-built robot image dataset, a multimodal robot control knowledge graph was constructed. Finally, to address the limitations of conventional KGAT, a Personalized Knowledge Graph Attention Network (PKGAT) model is proposed by integrating the recommendation system with the constructed knowledge graph and incorporating a cold-start mitigation module. This results in a knowledge graph-based recommendation system tailored for robotic control domain knowledge. The results from the experiments show that the AUC value and F1 value of the PKGAT model in the MovieLens dataset are 94.56% and 94.20% respectively, which have increased by 2.35% and 2.07%, respectively, compared with the KGAT model. The AUC value and F1 value in the Book-Crossing dataset were 75.30% and 73.60%, respectively, which increased by 2.02% and 1.48%, respectively, compared with the KGAT model.