在個人化穿搭推薦的領域中，產生既時尚又受約束的目錄的推薦是一項挑戰。本論文提出一套名為〈運用外部穿搭錨點進行來源目錄推薦〉的混合式推薦系統，結合了基於內容的相似度匹配與規則式過濾策略。系統首先在來源目錄與外部時尚資料集中，找出語意與視覺上相似的服飾項目。接著，將這些匹配項目對應至外部資料集中的穿搭組合，這些組合被稱為「風格錨點」。系統再透過屬性相似度，將每個錨點中的構成元素投影回來源目錄中，從而生成完全基於原始目錄的風格一致之穿搭建議。最終推薦結果會透過特定領域知識的規則進行精煉，以強化其在季節性、色彩和諧性與服裝結構上的相容性。本研究結合了外部時尚知識的創造力與來源目錄限制下的實用性。系統在一個真實的男裝資料集上進行驗證，實驗結果顯示在多樣性、涵蓋率與規則遵從性方面，相較於基線皆有明顯提升。;In personalized outfit recommendation domain, generating recommendations that are both stylish and constrained catalog presents significant challenges. This thesis introduces a hybrid recommender system titled “Leveraging External Outfit Anchors for Source Catalog Recommendation”, which integrates content-based similarity matching with rule-based filtering. First, the system finds semantically and visually similar clothing items between a source catalog and an external fashion dataset. From these, the system maps matched items to outfit combinations from the external catalog, which are referred to as style anchors. Each component of these anchors is then projected back to the source catalog via attribute-based similarity, allowing the generation of a coherent outfit recommendation, solely within the original catalog. The final recommendations are refined using domain-specific rules to enhance compatibility in terms of season, color harmony, and garment structure. This work leverages the creativity of external fashion knowledge with the practicality of source catalog constraints. The system is validated on a real-world menswear dataset and the experimental results show improvements in diversity, coverage, rule adherence compared to baselines.
