From Knowledge Generation to Knowledge Co-creation: A Trust-Aware Recommendation System Based on Retrieval Augmented Generation

Abstract

In the era of information explosion, academic researchers are easily caught into the issue of “information overload”, which has attracted more and more attention. Researchers have an urgent need for obtaining scholarly information in their own fields efficiently and accurately, which makes recommendation system play an increasingly important role in scholarly information service. Most of the existing scholarly information recommendation systems adopt users' explicit input or implicit feedback for modeling and prediction, combined with deep learning mechanism to improve the recommendation performance. To capture the latent attributes of user-item interaction accurately, graph approaches, like knowledge graph, are widely utilized to represent information in a structured way. However, under the circumstance of big data, traditional recommendation strategies and methods exhibit significant limitations in semantic understanding and learning ability. Additionally, those approaches often neglect the critical role of social relations in shaping user’s interest and preference which have been proved and demonstrated in the context of knowledge recommendation. Therefore, traditional recommendation faces serious challenges to accurately understand user’s query, capture user’s interest during the recommendation process, and generate the accurate and reliable recommendations to meet user’s dynamically changing need. This inadequacy highlights the pressing need for the improved recommendation approach capable of integrating semantic insights, adaptive learning mechanisms for dynamically changing data, and social contextual factors to enhance recommendation performance in the scholarly context. Furthermore, traditional recommendation systems often struggle to adapt to dynamic and evolving knowledge landscapes. Static recommendation systems rely on static knowledge bases for generating recommendations, failing to account for the dynamic, diverse, and co-creative nature of knowledge. As a result, their recommendations often lack credibility and adaptability in ever-evolving collaborative environments.

To address the gaps, this study proposes a trust-aware recommendation system based on retrieval augmented generation (RAG) framework, utilizing modified RAG architecture as the recommendation framework to realize knowledge co-creation and taking scholarly trust relationship as a key factor in the knowledge co-creation processes. Unlike conventional recommendation systems that rely solely on pre-existing datasets or user-item interactions, our approach transforms the inherent paradigm of recommendation systems through leveraging RAG to retrieve and generate recommendations in real-time, ensuring adaptability to the continuous evolution of knowledge, and enabling a shift from knowledge generation to knowledge co-creation. Involving homophily theory and social influence theory, which have stated that a user’s preference is similar or influenced by their connected friends, this study suggests that the social relationship in academia leverages three primary components: co-authorship relationship, citation relationship, and social relationship. Therefore, one of the objectives in the study is to observe and identify the evolution patterns of social relationships in the scholarly context. Through extracting the co-authorship relationship from co-authorship network, the citation relationship from citation network and the social relationship from academic social network, this study combines these three types of scholarly relationships to construct the scholarly trust network and develops the metrics for trust value to quantify the strength of trust between researchers subsequently. In response to the research gaps identified in previous studies, a novel trust-aware recommendation approach based on RAG is adopted to enhance the recommendation relevance, accuracy and consistency. The RAG architecture combines a retriever, an augmenter and a generator to provide personalized recommendations. In the retrieval phase, the scholarly resources (e.g., academic articles) are chunked into vector embeddings and calculate the vector similarities to retrieve highly relevant academic documents. In the augmentation phase, a trust indicator is incorporated into the prompt and the generator employs a large language model (LLM) to summarize, contextualize, and explain the recommended content in the generation phase. This integration ensures that recommendations are not only accurate but also explainable and reliable, addressing a common limitation in existing recommendation systems.

To validate the feasibility and effectiveness of the proposed model, this study focuses on academic paper recommendation. Two real-world datasets are collected from ScholarMate.com, which is one of the largest academic social network platforms, used to construct the proposed recommendation framework and comparing the performance with three baseline recommendation models. Firstly, the paper dataset obtains the scholarly paper data and consists of the metadata of articles with the topics on “Artificial Intelligence + Recommendation Systems” including title, authors, abstract, journal, references, etc., which helps identify citation relationship and co-authorship, and construct citation network and co-authorship network respectively. Secondly, the user dataset obtains scholarly social network data and user behavioral data, containing the data about followings and followers, the data about “saved items” and “liked items”, etc., which helps identify social relationships and construct trust network.

This study contributes to the literature of recommendation systems in three ways. First, this study transforms the inherent paradigm of recommendation systems through leveraging RAG and realizes the shift from knowledge generation to knowledge co-creation between knowledge contributors. Second, this study integrates the scholarly trust as a central factor in the recommendation process, which plays a crucial role in knowledge dissemination and co-creation. In this study, scholarly trust is composed of three dimensions: interpersonal trust which reflects the trust in the expertise of contributors, intellectual trust reflecting the trust in the credibility of retrieved knowledge, and collaborative trust indicating the trust in the effectiveness of the co-creation process. By embedding these trust dimensions into the RAG framework, the system ensures that recommendations are not only relevant but also credible and contextually aligned with the needs of the users. Third, the study develops a trust-aware recommendation system for scholarly information, which continuously monitors ongoing knowledge interactions, updates its retrieval mechanisms accordingly, and adjusts recommendations in real-time based on trust-weighted inputs. This ensures that users receive the most up-to-date, reliable, and context-aware recommendations, fostering a more interactive and participatory knowledge-sharing ecosystem. By integrating retrieval-augmented generation, trust modeling, and dynamic adaptation, this research advances the field of intelligent recommendations in knowledge-intensive environments. The proposed recommendation system enables users to engage in a more interactive, trust-aware, and continuously evolving recommendation process, facilitating more effective and meaningful knowledge co-creation. This work has significant implications for academia, industry, and digital knowledge-sharing communities, paving the way for more robust and adaptive recommendation systems in the future. Future work will focus on refining scholarly trust models, expanding the recommendation system’s applicability to interdisciplinary research, and exploring the ethical implications of automated scholarly recommendations.

Date of Award	29 Sept 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Jian MA (Supervisor)