基於多向量融合與 Transformer 用戶偏好建模的個性化新聞推薦;Personalized News Recommendation via Multi-Vector Fusion and Transformer-based User Preference Modeling

NCU Institutional Repository > 管理學院 > 資訊管理研究所 > 博碩士論文 > Item 987654321/98263

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98263

题名:	基於多向量融合與 Transformer 用戶偏好建模的個性化新聞推薦;Personalized News Recommendation via Multi-Vector Fusion and Transformer-based User Preference Modeling
作者:	吳昇航;Wu, Sheng-Hang
贡献者:	資訊管理學系
关键词:	新聞推薦;多向量融合;個性化推薦;Transformer;對比學習;用戶建模;News Recommendation;Personalized Recommendation;Multi-vector Fusion;Transformer;User Preference Modeling;Contrastive Learning
日期:	2025-07-09
上传时间:	2025-10-17 12:33:37 (UTC+8)
出版者:	國立中央大學
摘要:	在數位資訊爆炸的時代，新聞推薦系統扮演著協助用戶過濾資訊與提升閱讀體驗的關鍵角色。然而，現有方法在新新聞推薦情境中仍面臨冷啟動、語義理解不足及內容同質化等挑戰。傳統內容過濾、協同過濾與知識圖譜技術雖具一定效果，卻難以同時兼顧即時性、多樣性與語義精準度。為解決上述問題，本文提出一套結合大型語言模型（LLM）與多模態特徵融合的新聞推薦架構，將每篇新聞分解為三種語意向量：由圖片推理視覺吸引力的「吸引力向量」、透過對比學習對齊摘要與 LLM 生成全文的「內容向量」，以及推理多層次分類標籤的「支援向量」。三向量融合後，輸入單層 Transformer 結合用戶歷史行為進行建模與偏好學習，生成個性化推薦結果。實驗以 MIND-large 資料集為基礎，與 DFM、DKN、NAML、NRMS、LSTUR 及GNewsRec 等主流模型進行比較，結果顯示本模型在 nDCG、Hit@K、MRR 與 AUC 等各項指標上皆具顯著優勢，特別在處理冷啟動新聞上表現突出。進一步的敏感度與消融實驗亦驗證三向量設計具良好穩定性與模組互補性。綜上，本文的主要貢獻包括：（1）首度結合視覺心理學與 LLM 進行圖片吸引力建模；（2）融合生成式語言能力與對比學習以強化新聞語義表徵；（3）透過語義分類提升推薦準確性與多樣性；（4）整合多模態表示與 Transformer 用戶建模，有效因應冷啟動與同質化挑戰。研究結果展現 LLM 在智慧新聞推薦中整合語義、視覺與語境資訊的潛力，具高度學術與應用價值。 ;In the era of digital information overload, news recommendation systems play a crucial role in helping users filter vast amounts of information and enhancing their reading experience.However, existing news recommendation methods still face several challenges, particularly in recommending newly published news items, including cold-start issues, insufficient semantic understanding, and recommendation homogeneity. Traditional approaches based on content filtering, collaborative filtering, or knowledge graphs offer partial solutions, yet often fall short in simultaneously addressing the timeliness, diversity, and semantic precision required by modern news recommendation. To overcome these limitations, this study proposes a novel news recommendation framework that combines Large Language Models (LLMs) with multi-modal feature fusion. The core innovation lies in decomposing news information into three distinct vector modules:(1) the Attractive Vector, derived from visual analysis of news images using LLMs to estimate their visual salience and psychological appeal;(2) the Content Vector, enhanced through contrastive learning by aligning summaries with full-length news content generated by LLMs to capture deep semantic meaning; and(3) the Supporting Vector, constructed by inferring multi-level news category labels via LLMs to enrich semantic recognition. These three vectors are fused and subsequently fed into a Transformer model to capture user reading behavior. The resulting user embedding is then combined with candidate news vectors to generate personalized recommendation results. Experiments are conducted on the MIND-large dataset, where news articles are paired with annotated images. The proposed model is compared against several mainstream baselines, including DFM, DKN, NAML, NRMS, LSTUR, and GNewsRec. Evaluation metrics include nDCG, Hit@K, MRR, and AUC. Results demonstrate that the proposed model outperforms all baseline methods across all evaluation criteria, with particularly notable advantages in new item recommendation. Furthermore, sensitivity analyses and ablation studies confirm the complementarity and stability of the proposed tri-vector architecture. In summary, the contributions of this work include:(1) Introducing LLMs to assess image attractiveness based on visual psychology for use in news recommendation;(2) Enhancing semantic representation of news content by combining generative language modeling with contrastive learning;(3) Improving recommendation precision and diversity through semantic category inference; and(4) Addressing cold-start and content redundancy issues via multimodal fusion and Transformer-based user modeling. This study demonstrates the potential of large language models in unifying semantic, visual, and categorical knowledge within multimodal recommendation systems, offering significant academic and practical value for the development of next-generation intelligent news recommendation systems.
显示于类别:	[資訊管理研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	16	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....