基于知识图谱嵌入的多模型融合与多任务学习的工业软件组件推荐方法
Multi-Model Fusion and Multi-Task Learning Based on Knowledge Graph Embedding for Industrial Software Component Recommendation

随着物联网、云计算、工业互联网等的快速发展，工业软件系统逐渐向扁平化、松耦合、平台化、模块化与组件化的架构模式发展。工业软件组件已逐步演变为跨域、跨空间、跨平台的网络化协同新模式。然而，由于工业软件组件数量的不断的增加，为创建工业软件开发挑选合适的工业软件组件变得更加困难。为了解决这个问题，已经提出了各种方法来推荐工业软件组件以匹配子任务的需求。但目前的方法在特征融合与利用、文本需求理解、子任务与组件的类别等元数据信息利用等方面存在一些挑战。为了克服现有服务推荐模型推荐效率不高以及学习能力不强的问题，本文提出了一种新的基于知识图谱嵌入的多模型融合与多任务学习的工业软件组件推荐模型(KG-MTFM)。模型利用语义特征提取模块来提取子任务需求与工业软件组件描述文档的语义特征，获得语义特征向量。建立组件分类的知识图谱通过表示学习算法获得实体嵌入向量，将其与组件语义特征向量链接形成工业软件组件的特征向量后，引入特征交互模块来对子任务和工业软件组件之间的特征交互进行建模。实验结果表明，我们的方法优于目前流行的主要方法。
With the rapid development of the Internet of Things, cloud computing, industrial Internet, etc., industrial software systems are gradually developing into flat, loosely coupled, platform-based, modular and component-based architecture models. Industrial software components have gradu-ally evolved into a new mode of cross-domain, cross-space and cross-platform network collabora-tion. However, due to the continuous increase in the number of industrial software components, it becomes more difficult to select the right industrial software components for creating industrial software development. To solve this problem, various methods have been proposed to recommend industrial software components to match the needs of subtasks. However, the current methods have some challenges in the aspects of feature fusion and utilization, text requirement understanding, subtask and component classification and other metadata information utilization. In order to over-come the problems of low recommendation efficiency and weak learning ability of existing service recommendation models, this paper proposes a new industrial software component recommenda-tion model (KG-MTFM) based on multi-model fusion and multi-task learning embedded in knowledge graph. The model uses the semantic feature extraction module to extract the semantic features of subtask requirements and industrial software component description documents, and obtain the semantic feature vector. The entity embedding vector is obtained by the representation learning algorithm, which is linked with the component semantic feature vector to form the feature vector of the industrial software component, and the feature interaction module is introduced to model the feature interaction between the subtask and the industrial software component. The ex-perimental results show that our method is superior to the prevailing methods.