A Human-AI Collaboration-Enhanced Recommendation Framework for Patent Co-Development in University

Abstract

Industrial alliances—regional clusters of related firms—have become a key driver of innovation. While innovation is essential for competitiveness, many alliances in developing countries struggle due to limited funding and rely on government-supported R&D infrastructure. However, this infrastructure cannot always keep up with fast-changing technological needs. Prior research shows that firms can reduce R&D costs and time by using academic patents. This thesis proposes a novel model to identify suitable common academic patents for industrial alliances, helping them innovate more efficiently. Two main challenges in applying this model are also identified. First, the vast volume of academic patents creates a severe information asymmetry between universities and industry alliances. Second, although existing patent recommendation methods can suggest relevant patents to companies, the companies still need to allocate time to communicate with multiple busy scholars who invented these patents and decide which patents to purchase from the recommendation list. This inefficient communication process impedes the transfer of academic patents and reduces the likelihood of subsequent patent co-development between universities and industry alliances. To address these challenges, this thesis proposes a human-AI collaboration-enhanced recommendation framework. This framework consists of two components: the first is an attentive multi-LSTM-CapsNet method designed to identify common technical needs within industry alliances and complete patent recommendations; the second is a human-multi-agent collaborative framework with a resource-based view communication template to improve the efficiency of patent purchase decisions and assist companies in identifying potential opportunities for patent co-development.

The first component involves recommending patents that meet the common needs of industry alliances. Previous patent recommendation methods primarily focus on addressing the technical needs of individual companies and fail to effectively meet the common technical requirements of all companies within an industry alliance. Moreover, previous approaches often concentrated solely on analyzing historical patent sequences to understand a company's past technological preferences, overlooking potential technological opportunities. This includes factors like the influence of domain knowledge on the future technological direction of industry alliances, such as the development planning of key common technologies in the industry. To address the problem, this study proposes a deep learning (DL) method that recommends patents to industrial alliances based on common technological needs mining. The proposed method mines the common needs from patents owned by the companies and domain knowledge about potential technologies common to alliances. Specifically, we mine the technological needs of the companies from their patents using long short-term memory (LSTM) networks and obtain their patent-based common needs by designing a candidate patent-aware attention mechanism. Then, we extract implicit technology directions from the domain knowledge using a capsule network (CapsNet) and obtain domain knowledge-based common needs by designing an industrial alliance-aware attention mechanism. We evaluate the proposed method through offline experiment, comparing it to various benchmark methods. The experimental results demonstrate that our method outperforms these benchmarks in terms of recall and normalized discounted cumulative gain. To the best of our knowledge, this is the first study that recommends academic patents to industrial alliances using a novel DL method based on common technological needs mining. The methodology innovation is attentive multi-LSTM-CapsNet for common technological need mining.

The second component focuses on enhancing communication and collaboration between alliance companies and digital scholars. Existing research has primarily concentrated on designing specific digital scholars for teaching and research contexts, typically utilizing a one-on-one interaction mechanism where students or scholars engage with digital scholars. However, in the context of academic patent transfer and co-development between universities and industry alliances, companies need to engage with multiple personalized scholars, which involves a one-to-many interaction model, and the content of these interactions differs significantly from the teaching and research scenarios. To address this challenge, this study proposes a company-multi-digital scholar collaborative framework, incorporating a resource-based view communication template to improve the efficiency of patent purchase decisions and help companies identify potential opportunities for patent co-development. Specifically, we employed a Large Language Model-based (LLM-based) agent framework to design digital scholars. To tailor the digital scholars to the scenario of this research, we first used the Resource-based View (RBV) theory to identify 11 key factors, which not only scholars can provide but also influence a company’s patent purchase decisions and patent co-development opportunity discovery. These factors include patent price, novelty, quality, maturity, license exclusivity, inventor reputation, applied research background, university reputation, technology transfer experience, prior collaboration with companies on projects, and analysis of technology opportunities for target patents. We then defined these 11 factors as tools and embedded them into the LLM-based agent framework's tool module. Finally, we developed a communication template for digital scholars, guiding companies in their interactions around the aforementioned key concepts. This approach enables companies to enhance information acquisition efficiency by communicating with digital scholars representing multiple recommended patent inventors. We conducted a field experiment to evaluate the effectiveness of the proposed framework, using two indicators: the number of digital scholars the company interacts with and the frequency of interactions between the company and digital scholars. This study represents the first effort to promote patent transfer and co-development between universities and industry alliances through interactions between companies and multiple digital scholars. The methodological innovation lies in the human-multi-agent collaborative framework with a resource-based view communication template.

This thesis presents two innovative methods: the first is an attentive multi-LSTM-CapsNet approach for mining common technical needs, and the second is a human-multi-agent collaborative framework with a resource-based view communication template, designed to improve the efficiency of information acquisition for companies in academic patent purchase decisions and capture opportunities for patent co-development. The theoretical contribution lies in extending the RBV theory to technology transfer scenarios and applying it to design interaction templates between companies and digital scholars. The practical contribution lies in establishing a dynamic, company-oriented patent co-development environment between universities and industry alliances. For universities, this framework enhances the commercialization of academic patents. For industry alliances, it improves technology identification efficiency, shortens innovation cycles, and boosts the overall effectiveness of technological innovation.

Date of Award	4 Jun 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Jian MA (Supervisor)