Classification Analysis And Structure Pattern Detection On Patent Citation Network

Patent analysis mainly covers two important aspects: patent classification analysis and prediction of patent technology trends. The traditional method of classification analysis only concerned node’s attribution or topological structure, but it didn’t completely get enough information from patent citation network. Meanwhile, the research on predicting the trends of patent technology from patent citation network focused on the direct citation of a single patent node, so the domain knowledge flow could not be parsed. To solve the above problem, this thesis synthesizes node attribution, node characteristic of topological structure and structural pattern, puts forward method of supervised learning on patent classification analysis, and on this basis proposes detecting structural pattern for predicting the trends of patent technology. The main innovative research work of this thesis can be summarized as follows:1. Modeling and centrality analyzing of patent citation network. It introduces relationship of patent citation, especially the concept of patent element, citing literature, and the categories of patent citation network. Then a new specific framework and method is employed for the centrality of patent citation network.2. Adopting the graph neural network in patent classification analysis. It exploits the projection methodology in connection with patent classification analysis, by projecting the patent node to hyperspace for data analysis. Based on the model of graph neural network, it synthesizes node attribution, node characteristic of topological structure and structural pattern, and accomplishes the projection from network analysis to data analysis in hyperspace by supervised learning. In addition, it concretely presents how to carry out the learning of model parameter and implementing of the model. Finally, it takes artificial network and actual network as examples to prove the effectiveness and feasibility of this method.3. Revealing the process of discovering structural pattern about patent citation network. The thesis gives its definition firstly, represents it by thumbnails and mapping matrixes, and proposes algorithm of discovering structural pattern on the basis of density. Furthermore, it proposes the hypothesis that ideal structural patterns play an important role in analyzing network structure and grasping the method of knowledge flow, so it converts the discovery of structural pattern to graph matching, and adopting the method of graph simulation in mining and decomposing the ideal structural pattern from patent citation network. With the analysis of centrality, this thesis proves that it is more scientific and accurate to detect the central patent by taking advantage of structural pattern of the patent citation network.