本研究探討如何將大型語言模型（Large Language Models, LLMs）的語意推理能力有效融入圖神經網路（Graph Neural Networks, GNNs）推薦系統，同時避免既有方法所面臨的大量語意表示生成、高維嵌入對齊及推論成本過高等問題。現有 LLM-enhanced recommendation 方法多以節點層級的語意表示為核心，透過生成使用者或物品描述、語意嵌入對齊或圖結構增強等方式提升推薦效能，但較少考慮多跳協同路徑所蘊含的語意關聯，且往往需要大量 LLM 推論與跨模態表示學習。

為解決上述限制，本研究提出 Subgraph Semantic Learning (SSL)，以證據子圖（evidence subgraphs）作為語意知識注入的基本單位。首先，自使用者與候選物品之間擷取具代表性的協同路徑，並透過分群建立多個證據子圖；接著利用 LLM 執行遮罩重建（masked reconstruction）任務，評估各子圖所包含協同證據的語意合理性，再將推理結果轉換為輕量化的語意分數（semantic scores）。最後，以多任務學習方式將此語意監督導入圖推薦模型，使模型在維持協同過濾能力的同時，學習將較高注意力分配給具有較高語意合理性的證據子圖，而無須進行高維文字嵌入對齊。

本研究於 MovieLens-1M、Netflix 與 Amazon-Books 三個公開推薦資料集上進行實驗，並結合 LightGCN、GCCF、NGCF 與 SimGCL 等多種圖推薦模型進行驗證。實驗結果顯示，SSL 能穩定提升不同 GNN backbone 的推薦效能，並在多項評估指標上達到具競爭力的表現。消融實驗亦證實，相較於固定分數或隨機分數，LLM 所產生的語意分數能提供更有效的監督訊號；此外，僅需對少量使用者進行 LLM 標註，即可獲得明顯的效能提升，展現良好的資料效率與可擴展性。由於 LLM 僅於離線訓練階段使用，推論階段完全不需依賴 LLM，因此能兼顧推薦準確性、可解釋性與實際部署效率。

Graph Neural Networks (GNNs) have achieved remarkable success in recommendation systems by effectively modeling high-order collaborative filtering signals from user–item interaction graphs. However, conventional GNN recommenders rely primarily on interaction topology and often fail to distinguish between structurally similar but semantically different collaborative evidence. Meanwhile, recent Large Language Model (LLM)-enhanced recommendation approaches typically inject semantic knowledge through node-level representations, dense embedding alignment, or graph augmentation, resulting in substantial annotation costs, cross-modal optimization difficulties, and limited exploitation of multi-hop collaborative reasoning. This thesis therefore aims to make evidence subgraphs “speak” by endowing them with semantic reasoning capabilities provided by LLMs.

To achieve this goal, this thesis proposes Subgraph Semantic Learning (SSL), a novel framework that introduces semantic supervision at the evidence-subgraph level. Instead of generating semantic representations for all users and items, SSL first retrieves representative collaborative paths between a target user–item pair and organizes them into compact evidence subgraphs. An LLM then performs a masked reconstruction task to evaluate the semantic plausibility of each subgraph. The reconstruction quality is transformed into lightweight scalar semantic scores, which are subsequently used to supervise the attention mechanism of a graph recommender through a multi-task learning objective. By representing semantic knowledge as scalar supervision rather than high-dimensional textual embeddings, SSL effectively avoids cross-modal representation alignment while preserving the reasoning capability provided by LLMs.

Experiments are conducted on three benchmark datasets, namely MovieLens-1M, Netflix, and Amazon-Books, using multiple GNN backbones including LightGCN, GCCF, NGCF, and SimGCL. Experimental results demonstrate that SSL consistently improves recommendation performance across different graph recommendation architectures while achieving competitive results against recent LLM-enhanced recommendation methods. Ablation studies further confirm that the proposed LLM-derived semantic scores provide substantially more informative supervision than constant or random alternatives. In addition, meaningful performance improvements can be obtained with semantic annotations from only a small subset of users, demonstrating the data efficiency and scalability of the proposed framework. Since the LLM is required only during offline training, the final recommendation model introduces no inference-time LLM overhead, making SSL an efficient and practical solution for integrating semantic reasoning into graph-based recommendation systems.

[1] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016.
[2] Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl. Item-based collaborative filtering recommendation algorithms. In Proceedings of the 10th international conference on World Wide Web, pages 285–295, 2001.
[3] Yehuda Koren, Robert Bell, and Chris Volinsky. Matrix factorization techniques for recommender systems. Computer, 42(8):30–37, 2009.
[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020.
[5] Ziqiang Cui, Yunpeng Weng, Xing Tang, Fuyuan Lyu, Dugang Liu, Xiuqiang He, and Chen Ma. Comprehending knowledge graphs with large language models for recommender systems. In Proceedings of the 48th international ACM SIGIR conference on research and development in information retrieval, pages 1229–1239, 2025.
[6] Rich Caruana. Multitask learning. Machine learning, 28(1):41–75, 1997.
[7] Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. Neural graph collaborative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval, pages 165–174, 2019.
[8] Lei Chen, Le Wu, Richang Hong, Kun Zhang, and Meng Wang. Revisiting graph based collaborative filtering: A linear residual graph convolutional network approach. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 27–34, 2020.
[9] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 639–648, 2020.
[10] Junliang Yu, Hongzhi Yin, Xin Xia, Tong Chen, Lizhen Cui, and Quoc Viet Hung Nguyen. Are graph augmentations necessary? simple graph contrastive learning for recommendation. In Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval, pages 1294–1303, 2022.
[11] Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. Llmrec: Large language models with graph augmentation for recommendation. In Proceedings of the 17th ACM international conference on web search and data mining, pages 806–815, 2024.
[12] Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. Representation learning with large language models for recommendation. In Proceedings of the ACM web conference 2024, pages 3464– 3475, 2024.
[13] Xinji Zha, Yumin Dong, Haomiao Jiang, Zihui Xu, and Chao Wang. Llm-agr: Large language model augmented graph representation learning for recommendation. Knowledge-Based Systems, page 114791, 2025.
[14] Leheng Sheng, An Zhang, Yi Zhang, Yuxin Chen, Xiang Wang, and Tat-Seng Chua. Language representations can be what recommenders need: Findings and potentials. In international Conference on Learning Representations, volume 2025, pages 91632–91658, 2025.
[15] Yu Wang, Lei Sang, Yi Zhang, and Yiwen Zhang. Intent representation learning with large language model for recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1870–1879, 2025.
[16] Gobinda G Chowdhury. Introduction to modern information retrieval. Facet publishing, 2010.
[17] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. Introduction to Algorithms. MIT press, 3rd edition, 2009.
[18] James B McQueen. Some methods of classification and analysis of multivariate observations. In Proc. of 5th Berkeley Symposium on Math. Stat. and Prob., pages 281–297, 1967.
[19] Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006.
[20] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017.
[21] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 3982–3992, 2019.
[22] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.
[23] Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1):79–86, 1951.
[24] F Maxwell Harper and Joseph A Konstan. The movielens datasets: History and context. Acm transactions on interactive intelligent systems (tiis), 5(4):1–19, 2015.
[25] James Bennett, Stan Lanning, et al. The netflix prize. In Proceedings of KDD cup and workshop, volume 2007, page 35, 2007.
[26] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, pages 43–52, 2015.
[27] Kalervo J¨arvelin and Jaana Kek¨al¨ainen. Cumulated gain-based evaluation of ir techniques. ACM Transactions on Information Systems (TOIS), 20(4):422–446, 2002.
[28] D Manning Christopher, Raghavan Prabhakar, and Schutze Hinrich. Introduction to information retrieval, 2008.
[29] EM Voorhees. Proceedings of the 8th text retrieval conference. TREC-8 Question Answering Track Report, pages 77–82, 1999.
[30] Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt- Thieme. Bpr: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618, 2012.
[31] Meta AI. Llama 3 model card. https://github.com/meta-llama/llama3, 2024.
[32] Jules White, Quchen Fu, Sam Hays, Michael Sandborn, Carlos Olea, Henry Gilbert, Ashraf Elnashar, Jesse Spencer-Smith, and Douglas C Schmidt. A prompt pattern catalog to enhance prompt engineering with chatgpt. arXiv preprint arXiv:2302.11382, 2023.
[33] Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. Minilm: Deep self-attention distillation for task-agnostic compression of pretrained transformers. Advances in neural information processing systems, 33:5776–5788, 2020.