Adaptive Taxonomy Learning and Historical Patterns Modeling for Patent Classification

• DOI: 10.1145/3674834
• Publication Date: 2024-07-11T16:24:28Z
• Authors: Tao Zou; Le Yu; Junchen Ye; Leilei Sun; Bowen Du; Deqing Wang
• Keywords: Leverage (statistics)

Patent classification aims to assign multiple International Classification (IPC) codes a given patent. Existing methods for automated patent primarily focus on analyzing the text descriptions of patents. However, apart from textual information, each is also associated with some assignees, and knowledge their previously applied patents can often be valuable accurate classification. Furthermore, hierarchical taxonomy defined by IPC system provides crucial contextual information enables models leverage correlations between improved accuracy. existing fail incorporate above aspects lead reduced performance. To address these limitations, we propose an integrated framework that comprehensively considers patent-related specific, first present learning module capture both horizontal vertical within codes. This effectively captures adaptively exchanging aggregating messages among at same level (horizontal information) parent children (vertical information), which allows comprehensive integration relationships taxonomy. Additionally, design historical application patterns component previous corresponding assignee high-order temporal via dual-channel graph neural network. Finally, our approach combines texts, encompasses semantics codes, assignees’ sequential preferences make predictions. Experimental evaluations real-world datasets demonstrate superiority proposed over methods. Moreover, model’s ability assignees semantic dependencies