隨著影片資料在社群影音、線上教育、智慧監控與智慧城市等應用中的快速成長，影片表徵學習已成為影片理解中的重要研究議題。相較於傳統單一類別影片分類，多標籤影片分類更能描述真實影片中同時包含物件、場景、動作、事件與屬性等多層次語意的特性。近年來，對比學習（Contrastive Learning， CL）已廣泛應用於表徵學習，其目標是拉近語意相似樣本並推遠語意不同樣本。然而，多標籤影片之間常存在部分語意重疊，使傳統正樣本與負樣本的二元關係難以完整描述細緻的語意關聯。為解決上述問題，本研究提出 Medoid-Guided Multi-Label Contrastive Learning with Swapping Reconstruction（MMRCL），一個針對多標籤影片表徵學習所設計的結構化對比學習框架。MMRCL 由 V2P、Swapping、V2V 與 P2M 四個模組組成，分別用於全域語意對齊、共同與差異語意分解、分支層級關係排序，以及原型與代表性樣本校準。本研究於 HVU mini 與 Charades 兩個多標籤影片資料集上進行實驗，結果顯示 MMRCL 相較於多數基準方法取得較佳表現。進一步的表示空間分析、消融實驗、參數敏感度分析與案例研究亦驗證 MMRCL 能學習更符合多標籤語意關係的影片表示，並展現各模組的互補效果。;With the rapid growth of video data in social media， online education， smart surveillance， and smart city applications， video representation learning has become an important research topic in video understanding. Compared with conventional single-label video classification， multi-label video classification better describes real-world videos containing objects， scenes， actions， events， and attributes simultaneously. Recently， Contrastive Learning (CL) has been widely used in representation learning to pull semantically similar samples closer and push dissimilar samples apart. However， multi-label videos often contain partial semantic overlap， making binary positive and negative relationships insufficient for modeling fine-grained semantic relations. To address this issue， this study proposes Medoid-Guided Multi-Label Contrastive Learning with Swapping Reconstruction (MMRCL)， a structured contrastive learning framework for multi-label video representation learning. MMRCL consists of four modules， V2P， Swapping， V2V， and P2M， which respectively perform global semantic alignment， shared and distinct semantic decomposition， branch-level relational ranking， and prototype-to-medoid calibration. Experiments on HVU mini and Charades show that MMRCL outperforms several baseline methods. Further representation space analysis， ablation studies， sensitivity analysis， and case studies verify that MMRCL learns video representations that better reflect multi-label semantic relationships and demonstrate the complementary effects of the proposed modules.
