HE Xijun, SHI Anjie, WU Shuangshuang, WU Yuying
Patent transactions play an important role in improving technological innovation capacity of the demand side and the radiation capacity of the supply side. However, the number of patents in China's technology market is large, but the transaction rate between patents is still at a low level. Therefore, exploring paths of promoting patent transaction rate receives plenty of attention. In this paper, we select multidimensional factors affecting patent transactions based on former studies. Then, the semantic and non-semantic entities in patent information, as well as the multi-dimensional relationships between the entities, are used to construct the patent supply and demand knowledge graph (PSD-KG). Meanwhile, TransE model is used for knowledge graph embedding representation. On this basis, reinforcement learning is introduced to construct a patent transaction recommendation model based on knowledge graph and reinforcement learning (KG-RL-PTR). Then, the reward functions are designed based on the entity similarity relationship and the organization's historical transaction information, which guides the agent to reason out the effective paths based on the environment in which the organization is located on the PSD-KG, so as to search for potential high-quality patents and complete the transaction recommendation. An empirical study is conducted based on patent data in the field of fuel cells. We find that: 1) The KG-RL-PTR model has the most superior recommendation performance when compared with Ekar, DDPG, DeepPath and other methods. 2) Inter-organizational technological proximity is the key path for patent transaction recommendation. Meanwhile, the paths based on social proximity, geographic proximity, and institutional proximity also contribute to the recommendation process. 3) The interaction between multidimensional proximities jointly affects the formation of patent transaction relationships. Among them, technological proximity and social proximity play dominant roles, which suggests that similarity between technologies and inter-organizational trust significantly affect patent recommendation. Their interactions with other multidimensional proximities, including institutional proximity and geographic proximity, jointly influence recommendation results. This model provides inference paths while giving recommendation results, and it improves the accuracy, novelty and interpretability of recommendation results.
