Abstract: To solve the low-efficacy problem in leveraging universal recommendation algorithms for knowledge supply,a Context-aware GAN-based Knowledge Recommendation method (CGKR) was proposed to provide engineers with the proper knowledge.By considering the content similarity and task similarity,a knowledge relevance network was constructed for user historical ratings' semantic extending to fully explore user's interests.Based on the users' personal background information and historical behavior information,the background context and task context were constructed to get the user representation vectors contextually-detailed.To obtain the high recommendation efficacy,the Generative Adversarial Network (GAN) model was introduced,and the Context-aware GAN (CxtGAN) model using the user representation vectors with context information as collaborative training dataset was constructed to solve the sparse matrix and cold-start problems.An engineering case study was undertaken to illustrate the CGKR's effectiveness.As evidenced in the evaluation,CGKR was ensured the effectiveness of the knowledge recommended for target user in a specific task situation.

Key words: knowledge recommendation, generative adversarial network, context aware, spreading activation model, semantic network, knowledge management

摘要： 针对现有知识推荐方法因稀疏矩阵和冷启动导致推荐性能不佳的问题,提出一种基于情境感知生成对抗网络模型的知识推荐方法(CGKR)。提出任务相似度概念,同时考虑内容相似度和任务相似度构建知识相关性网络,基于知识相关性网络构建语义激活扩散模型,扩展用户历史评分,以全面探知用户兴趣;基于用户个人背景信息和历史行为信息构造用户情境和任务情境;引入生成对抗网络模型,并结合情境信息构建情境感知生成对抗网络模型(CxtGAN);基于训练完成的CxtGAN,为特定任务情境下的目标用户提供个性化知识推荐服务。以某船厂知识管理系统数据为例,进行实例分析与实验研究,结果表明CGKR方法具有较好的知识推荐性能,能够为企业用户提供优质知识推荐服务。

关键词: 知识推荐, 生成对抗网络, 情境感知, 激活扩散模型, 语义网络, 知识管理