Abstract: Organized industry-university-research (IUR) collaboration is crucial for leveraging China's new national system advantages, enhancing the overall efficacy of the national innovation system, and achieving deep integration of industry, university, and research. Identifying collaboration topics is paramount for high-quality, efficient synergy in such collaborations and constitutes a primary challenge. This paper proposes a set of identification methods for IUR collaboration topic pairs based on complex network analysis and deep learning algorithms. Firstly, this paper analyzes the concept of "Organized IUR Synergy Innovation" and related topic features, proposes that the topic of IUR collaboration should be high-value and strong-relevance. Secondly, a bi-layer network of "science keywords-universities and research institutions" and a bi-layer network of "technology keywords-enterprises" are constructed based on paper data and patent data. The SciBERT model is used to construct a semantic network of science keywords and a semantic network of technology keywords. The Node2vec-based link prediction model is used to generate the future semantic network of science keywords and the semantic network of technology keywords. Then, the five indicators of novelty, fundamentality, width, growth, and foresight are measured using methods of complex network topology analysis, community discovery, and machine learning, to identify high-value scientific and technological topics. Finally, the correlation between scientific and technological topics is calculated using the improved semantic similarity index, SimDoc. This enables the selection of “science-technology topic pairs” that exhibit high potential for collaboration across IUR. In this paper, an empirical study is conducted using paper and patent data in the field of “artificial intelligence (AI)” from 2018 to 2022 to validate the research method and results. This study identifies 20 high-value scientific and technological topics in the field of AI. The results reveal that scientific topics focused more on basic theoretical research such as neural network training algorithms, human-computer interaction methods, and large model training algorithms, while technological topics focused more on application research such as computer vision technologies, computing devices, and big data processing techniques. The correlation analysis results between scientific and technological topics show that the overall correlation is above 0.5, and there are 22 science-technology topic pairs with scores of 0.9 or above, indicating that organized IUR synergy innovation in the AI field has a good foundation. The primary theoretical contributions are as follows. First, integrating the concept of "organization" into the design of an efficient IUR collaborative system, deepening the theoretical research on IUR synergy innovation. Second, expanding the application of complex network analysis and deep learning algorithms in the field of IUR synergy innovation research. Third, the integration of machine learning and text semantic analysis methods provides a useful supplement for deep mining of the intrinsic relationships between IUR cooperation topics. Fourth, enriching the theoretical and methodological research in areas such as optimizing the allocation of innovative resources and transforming scientific and technological achievements. This study can provide important decision-making support for government departments to guide enterprises, universities, and research institutes to carry out high-level IUR collaborative innovation in an organized manner, and play an important quantitative supporting role in promoting the deep integration of the innovation chain and industrial chain in practice.

摘要： 开展有组织的产学研协同创新，是发挥我国新型举国体制优势、实现产学研深度融合的重要内容。其中，对产学研“合作主题”的有效识别是实现高质量高效率产学研协同创新的首要问题。本文提出了一套基于复杂网络分析和深度学习算法的产学研协同创新合作主题挖掘方法，首先，围绕“有组织的产学研协同创新”概念和主题特征进行深度剖析，提出产学研合作主题应具有高价值性和强相关性两大特征；其次，基于论文数据和专利数据分别构建“科学主题词-学研机构”双层网络和“技术主题词-企业”双层网络，其中，SciBERT模型被用来构建科学和技术主题词语义网络，基于Node2Vec的链路预测模型被用来预测未来的科学和技术主题词语义网络；之后，应用复杂网络拓扑结构分析、社区发现、机器学习等方法对主题的新颖性、基础性、广泛性、成长性、前瞻性五大指标进行测度，识别高价值的科学主题和技术主题；最后，对语义相似度指标SimDoc进行改进，计算科学主题和技术主题之间的相关性，遴选产学研协同潜力大的“科学主题-技术主题对”。本文选取人工智能领域开展实证研究，对提出的研究方法进行验证。本研究能为国家、区域和行业组织高层级产学研协同创新提供重要的量化决策参考。