18 March 2026 in IEEE Transactions on Computational Social Systems
Knowledge graphs (KGs) are a crucial foundation for intelligent systems performing advanced tasks such as personalized recommendations and question answering. However, many existing KGs are incomplete because of missing triples. Recent studies have shown that link prediction based on paths can be a reliable approach. Path-based methods mainly use path patterns within the KG to solve the puzzle of missing triples. These patterns, defined by the paths connecting two entities, are used to predict the relation between them. However, these methods are not perfect; relying solely on the relation between entities or their embedding information is ineffective in link prediction. To address the above challenges, we propose a new KG link prediction method leveraging the path and graph context. For a given candidate relation and two entities, the type hierarchy of entities is used for disambiguation while preserving the entities’ neighborhood information in the graph. Our method then encodes the path information using a bi-long short-term memory (LSTM), integrates the path patterns with the attention mechanism, and finally predicts whether the entities can be linked via the candidate relations. Experiments on three standard datasets demonstrate the effectiveness of our method.
27 March 2026 in Journal of Intelligent Manufacturing and Special Equipment
Purpose: To address the limitations of traditional failure mode and effects analysis (FMEA) methods in elevator fault analysis, including heavy reliance on human experience, limited use of large scale heterogeneous text data, static analysis results and insufficient interpretability, this study aims to develop an intelligent FMEA method for the elevator domain. Design/methodology/approach: An elevator FMEA method based on retrieval augmented generation (RAG) is proposed. An external knowledge base is constructed by integrating a knowledge graph (KG) with a vector database. During retrieval, a multi-route retrieval strategy is adopted to obtain candidate documents. A reranking model named CapsGCN-Rank based on a graph convolutional capsule neural network is designed to perform fine-grained filtering and reranking of candidate documents. The reranked documents are then combined with a large language model to generate structured fault analysis results. Findings: Experimental results show that the proposed method outperforms several baseline methods in context precision, context recall, as well as the relevance and correctness of generated answers. The method effectively improves the accuracy and completeness of elevator FMEA. Originality/value: The proposed approach introduces structured semantics from the KG, a multi-route retrieval strategy and dynamic routing of capsule neural networks into the RAG framework. It enables fine-grained document reranking and interpretable fault analysis for the elevator domain, providing an effective solution for FMEA in complex industrial scenarios.