序列推薦（Sequential Recommendation）旨在根據使用者的歷史互動行為，建模其偏好變化並預測下一個可能互動的項目。然而，在真實電子商務場景中，使用者互動紀錄往往呈現高度稀疏與長尾分布，使得純粹依賴 item ID 或行為序列的推薦模型難以有效學習穩定的項目表示。此外，現有序列推薦方法雖已廣泛採用 Transformer、圖神經網路或多模態特徵增強技術，但多數方法仍停留於淺層特徵整合，未能充分結合商品分類層級結構與深層語義資訊，因而限制了模型在冷啟動、跨類別泛化與推薦可解釋性上的表現。 為解決上述問題，本研究提出 TSSRec（Taxonomy-aware Semantic Sequential Recommendation），一種結合分類層級知識與大型語言模型語義增強的下一項推薦框架。本研究的核心概念在於，先透過分類結構嵌入與 LLM 語義嵌入之對齊，建立兼具結構性與語義性的統一項目表示空間，再透過序列推薦模型學習使用者的動態偏好狀態，並將其與候選項目表示進行匹配以完成下一項推薦。具體而言，本研究首先利用分類樹中的父子關係與兄弟關係學習項目的結構嵌入，使商品能保留其在 taxonomy path 中的層級位置與類別鄰近關係。接著，透過大型語言模型產生結構化語義摘要，並使用語義編碼器將商品標題、描述與多模態內容轉換為語義嵌入。為降低分類結構與語義表示之間的空間落差，本研究進一步設計跨模態對比學習機制，使相同項目的結構嵌入與語義嵌入在共享表示空間中相互靠近，進而形成統一且具解釋性的項目表示。最後，模型採用 SASRec 架構編碼使用者歷史行為序列，並以序列狀態與候選項目表示之內積分數完成 Top-K 推薦。 本研究於 Amazon Review Data 中兩個具代表性的電子商務資料集 Toys & Games 與 Industrial & Scientific 上進行實驗評估。實驗結果顯示，TSSRec 在 Hit Rate@K 與 NDCG@K 等標準排序指標上皆優於多個序列推薦與語義增強推薦基準模型。尤其在高度稀疏且專業性較強的 Industrial & Scientific 資料集中，TSSRec 能有效利用分類層級知識作為結構性錨點，並透過 LLM 語義摘要補足 item ID 無法表達的品牌、功能、用途與相容性資訊，進而提升長尾商品推薦的準確性與穩定性。消融實驗進一步驗證，分類結構嵌入、語義嵌入與對比學習模組皆對模型效能具有正向貢獻；個案分析亦顯示，本研究模型能透過 taxonomy path 與 LLM semantic keywords 提供具人類可理解性的推薦解釋。 本研究之主要貢獻包含：（1）提出一個分類感知語義對齊之序列推薦框架，透過對齊分類結構嵌入與 LLM 增強語義嵌入，建立統一且具解釋性的項目表示空間；（2）設計結合 Taxonomy分類層級知識與 LLM 生成的語義摘要的項目表示學習流程，以緩解資料稀疏與語義漂移問題；（3）透過實驗比較、消融分析、敏感度分析與個案研究，系統性驗證分類層級知識與語義增強資訊在下一項推薦任務中的有效性與可解釋性。;Sequential recommendation aims to model the evolution of user preferences from historical interaction sequences and predict the next item that a user is likely to interact with. However， in real-world e-commerce scenarios， user-item interactions are often highly sparse and long-tailed， making it difficult for models that rely solely on item IDs or behavioral sequences to learn robust item representations. Although recent sequential recommendation methods have incorporated Transformer architectures， graph neural networks， and multimodal feature enhancement， most of them still rely on shallow feature integration. They often fail to effectively combine the hierarchical structure of product taxonomies with deep semantic information， thereby limiting their performance in cold-start recommendation， cross-category generalization， and recommendation interpretability. To address these limitations， this study proposes TSSRec (Taxonomy-aware Semantic Sequential Recommendation)， a next-item recommendation framework that integrates taxonomic hierarchical knowledge with semantic enhancement from large language models. The core idea of TSSRec is to first construct a unified item representation space by aligning taxonomy-based structural embeddings with LLM-enhanced semantic embeddings. A sequential recommendation model is then used to learn dynamic user preference states and match them with candidate item representations for next-item prediction. Specifically， the proposed framework first learns structure embeddings from product taxonomies by modeling both parent-child relations and sibling relations， allowing each item to preserve its hierarchical position and categorical proximity along the taxonomy path. It then leverages a large language model to generate structured semantic summaries and employs a semantic encoder to transform product titles， descriptions， and multimodal content into semantic embeddings. To reduce the representation gap between taxonomic structure and semantic content， a cross-modal contrastive learning mechanism is introduced to pull the structural and semantic embeddings of the same item closer in a shared representation space. The aligned representations are then fused into a unified and interpretable item embedding. Finally， a SASRec-based sequence encoder is adopted to model the user’s historical behavior sequence， and Top-K recommendation is performed by computing the inner product between the learned sequence state and candidate item representations. Extensive experiments are conducted on two representative e-commerce datasets from Amazon Review Data， namely Toys & Games and Industrial & Scientific. The experimental results show that TSSRec consistently outperforms multiple sequential recommendation and semantic-enhanced recommendation baselines in terms of standard ranking metrics， including Hit Rate@K and NDCG@K. In particular， on the highly sparse and specialized Industrial & Scientific dataset， TSSRec effectively exploits taxonomic hierarchical knowledge as a structural anchor and uses LLM-based semantic summaries to capture brand， function， usage， and compatibility information that cannot be represented by item IDs alone. As a result， the proposed model improves both the accuracy and stability of long-tail item recommendation. The ablation study further confirms the positive contributions of taxonomy-based structural embeddings， semantic embeddings， and the contrastive alignment module. In addition， qualitative case studies demonstrate that TSSRec can provide human-understandable explanations through taxonomy paths and LLM semantic keywords. The main contributions of this study are threefold. First， we propose a taxonomy-aware semantic alignment framework for sequential recommendation， which aligns taxonomic structural embeddings with LLM-enhanced semantic embeddings to construct a unified and interpretable item representation space. Second， we design an item representation learning pipeline that combines taxonomy paths with LLM-generated structured summaries， thereby alleviating data sparsity and reducing semantic drift. Third， through performance comparison， ablation analysis， sensitivity analysis， and qualitative case studies， we systematically validate the effectiveness and interpretability of incorporating taxonomic hierarchical knowledge and semantic enhancement into next-item recommendation.
