Intelligent Patent Recommender Systems for Technology Transfer

Abstract

Technology transfer plays an increasingly important role in fostering product innovation in industry and improving national productivity. Through technology transfer activities, such as patent trading, companies can make full use of external technical resources to enhance their technological innovation ability and expedite product research and development. However, engaging in technology transfer activities require companies to invest a lot of manpower and time to identify appropriate technologies from a large number of potential selections. With the development of online technology transfer platforms and the accumulation of platform data, companies are facing an increasingly serious information overload problem.

Effective recommendation systems could alleviate the information overload problem and assist companies in timely discovery of appropriate patented technologies. This dissertation explores patent recommendations in three scenarios: recommendation for product innovation, requirement-oriented recommendation, and academic patent recommendation. In these scenarios, companies have different considerations regarding patents. This dissertation proposes a graph-based representation model to capture the latent characteristics of patents and companies, which exhibit complex relationships between them. Three novel approaches are developed to cater the special needs of companies.

Study 1 proposes a patent recommendation approach using multi-view learning for product innovation. The rapid changes in the technological composition of industries drive companies to seek patents from external sources for generating new product ideas. Current patent recommendation methods focus on analyzing the textual and bibliographic characteristics of users’ historical patents, but they ignore the technical trends of the company and its industry. To enable companies keep up with technology trends, this study takes multiple sources of content (i.e., industry categories, patent descriptions, and technology categories) into account and conducts multi-view representation learning for patents and target companies. A multi-layer classifier is trained to analyze patents at the industry level. Long short-term memory (LSTM) autoencoders are trained to embed the sequential graph-structured data that reflect technology trends into a continuous vector space. Experiment results show that the proposed approach obtains higher recommendation accuracy than the baseline approaches. Moreover, the multi-view recommendation approach outperforms the single-view approaches.

Study 2 develops a requirement-oriented patent recommendation approach using a technology-function matrix-based knowledge graph. With the development of online patent trading platforms, some companies proactively post their technology demands to seek patents from external sources. Although online patent marketplaces provide patents that can be traded, the knowledge gap between the requirements and available patents make it challenging to find the right patents that can meet specific requirements. This knowledge gap refers to the fact that companies’ requirements emphasize the functions/problems they want/face but lack other types of information, such as technical concepts and characteristics, which are usually the key information found in the patent texts. To bridge this knowledge gap, the study proposes a requirement-oriented patent recommendation approach. This approach builds a technology-function matrix-based knowledge graph to represent the technology-function details of patents and companies’ requirements. Finally, recommendation lists for supplementary, complementary, purely requirement-oriented, and hybrid recommendation strategies are obtained by comparing representation similarity. Experiments demonstrate the benefits of the constructed knowledge graph in identifying relevant patents that fulfill companies’ technology requirements.

Study 3 proposes a multi-dimensional trust-enhanced recommendation approach to promote academic patent transfers. Transferring technology from academia to industry has become a challenging task due to the “cultural divide” problem, where researchers in universities typically focus on knowledge discovery, while companies prioritize making profits through the application of proven technologies. This dichotomy creates a mistrust problem for companies when considering the use of academic patents developed by universities. To address this issue, this study proposes a multi-dimensional trust-enhanced recommendation approach aimed at facilitating academic patent transfers. The approach involves several steps: measuring tie strengths between companies and patents using a Personalized PageRank (PPR) model, assessing the trustworthiness of potential patent transactions by introducing a comprehensive trustworthiness measurement of academic patent, and integrating these measurements using a logistic regression model. User case studies demonstrate that the proposed recommendation approach yields more clicks and higher satisfaction scores compared to baseline approaches.

The main contributions of this dissertation include: (1) We design approaches specific to three patent recommendation scenarios: recommendation for product innovation, requirement-oriented recommendation, and academic patent recommendation. For each scenario, we identify the technical problems and provide corresponding solutions. (2) We introduce graph-based representation learning methods to represent features of patents, companies, and their interactions. In the recommendation scenario for product innovation, we employ LSTM model to represent the trends of companies’ interest and technology development. In the requirement-oriented recommendation scenario, we profile patents and companies based on the constructed knowledge graph. In the academic patent recommendation scenario, we develop the PPR model to represent the interaction features of company-patent interactions. (3) Our proposed approaches can be applied to online patent trading platforms as recommendation services to facilitate technology transfer.

Date of Award	19 Dec 2023
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Jian MA (Supervisor) & Yugang YU (External Supervisor)