离群专利多源特征视角下的颠覆性技术识别OACHSSCD

[Purpose/Significance]Disruptive technologies function as transformative forces with the capacity to rede-fine industrial structures and alter the trajectory of technological evolution.Their early identification is of strategic impor-tance for nations and organizations seeking to capture emerging opportunities and optimize S&T resource allocation.Cu-rrent research on disruptive technology identification faces two main limitations:overemphasis on mainstream technologies while overlooking outlier signals,and inadequate integration of multi-source features including semantic,metadata and relational characteristics.[Method/Process]In response to these limitations,we introduced a novel framework centered on outlier patents.This approach screened for technological outliers and utilized their multi-source characteristics to assess influence,thereby identifying disruptive technologies.Specifically,first,we adopted the SBERT model and the TF-IDF algorithm to construct the feature vectors of the patent text and the IPC classification number,respectively.After dimen-sion reduction,we concatenated them to obtain the composite patent vector.Second,we introduced three different outlier detection algorithms and adopted an incremental iterative strategy to select outlier patents with abnormal technical charac-teristics from the massive patent data.On this basis,we constructed a set of technical influence evaluation indicators.We also constructed the adjacency matrix according to the citation relationship of outlier patents,forming graph data including node characteristics and topology structure.Finally,we applied the Graph Convolutional Network(GCN)model.This model learned a representation vector that integrated the local graph context for each patent by iteratively propagating and aggregating the information of adjacent nodes,to accurately model the complex mapping relationship between patent fea-tures and technical influence labels.Ultimately,we achieved the prediction of disruptive technologies based on the multi-source characteristics of outlier patents.[Result/Conclusion]Based on a dataset of 17 577 brain-computer interface(BCI)patents retrieved from the Derwent Innovation platform(published before 2025)and employing advanced analytical tools implemented in Python(e.g.,PyTorch-based deep learning frameworks),our empirical analysis yields two key findings.First,the outlier-driven perspective provides a new paradigm for disruptive technology identification,shifting the focus from mainstream technological hotspots to marginal yet potentially transformative signals.This approach allows for the early detection of emerging technological trajectories that are typically overlooked by conventional indicator-based methods.Se-cond,the multi-feature influence model supports a more nuanced and comprehensive assessment of technological impact by integrating semantic(content-based),metadata(bibliometric),and relational(citation-based)dimensions,thereby over-coming the inherent constraints of single-dimensional metrics.These methodological innovations offer policymakers and R&D strategists a more sensitive and proactive approach for monitoring and evaluating emerging technological trends,with potential applications extending beyond the BCI domain to other fast-evolving fields of innovation.