短影片平台的快速興起，使推薦系統面臨內容高頻更新、模態間高度異質性、使用者行為時序衰減以及互動訊號分佈不均等多重挑戰。傳統協同過濾方法雖能利用交互資料學習偏好，但在稀疏或冷啟動場景下受限；純序列式模型擅長捕捉近期行為脈絡，卻難以描繪跨使用者的全域結構關聯；而多模態模型若缺乏與互動行為及語意層次的嚴格對齊，則難以在排序階段保持穩定表現。 基於此，本研究提出一個整合多模態內容與序列化互動的推薦框架，命名為SMCRec（Semantic-enhanced Multimodal Contrastive Recommendation）。SMCRec於單一體系下同時建模三條資訊路徑:1.語意路徑（Semantic）:利用 Gemini 生成影片語意摘要與主題標籤，並透過 OpenAI text-embedding-3-small 模型轉換為語意向量，強化語意層級的檢索能力；2.內容路徑（Content）:採用 ImageBind 將文本、視覺與聲音對齊至共享特徵空間，獲取魯棒的多模態內容表示；3.互動路徑（Interaction）:結合 LightGCN 捕捉長期結構性偏好，並整合序列模型 SASRec 吸收短期情境偏好，兼顧圖式擴散效應與近期序列動態。 三路徑嵌入經由獨立投影層映射至同維度空間，並採用跨路徑對比學習進行全方位對齊（涵蓋Semantic－Content、Semantic－Interaction 與 Interaction－Content），最後透過加權融合計算最終偏好分數。為優化模型訓練，本研究在互動路徑採用 BPR 損失，在對齊階段設計多工對比損失 (InfoNCE)，建構一致的語意空間。 實驗以 Short-video dataset（WWW 2025） 與 MicroLens 為主要資料來源，涵蓋短影片內容以及大規模、多型態的使用者互動行為。在評估指標 Hit@K、NDCG@K 與 MRR@K 上，本研究所提方法均取得整體提升。與現有的圖式推薦、多模態建模與序列式方法相比，本框架在語意可解釋性、跨模態一致性，以及長短期偏好的統一建模上，展現出系統化的改進；同時，其設計亦呼應近年研究趨勢，如圖神經網路中的鄰接聚合表徵學習與對比學習對表示空間對齊的正向效益。 本文的主要貢獻在於:提出一個三路徑 — 對比對齊 — 融合的一體化短影片推薦體系；將全域圖特徵注入序列化學習模型，兼顧長短期偏好；以 LLM 驅動的語意摘要縮短內容語意落差，並在無需昂貴標註的情況下維持良好可用性與穩定性;The rapid rise of short-video platforms has confronted recommendation systems with multiple challenges， including high-frequency content updates， high heterogeneity across modalities， temporal decay of user behaviors， and an uneven distribution of interaction signals. Although traditional collaborative filtering methods can leverage interaction data to learn preferences， they are limited in sparse or cold-start scenarios. Purely sequential models excel at capturing recent behavioral contexts， yet they struggle to delineate global structural correlations across users. Furthermore， if multimodal models lack strict alignment with interaction behaviors and semantic levels， they find it difficult to maintain stable performance during the ranking stage. To address these issues， this study proposes Semantic-enhanced Multimodal Contrastive Recommendation (SMCRec)， a recommendation framework that integrates multimodal content and sequential interactions. SMCRec simultaneously models three information paths within a unified architecture: 1.Semantic Path: Utilizes Gemini to generate video semantic summaries and hashtags， converting them into semantic vectors via the OpenAI text-embedding-3-small model to enhance retrieval capabilities at the semantic level. 2.Content Path: Employs ImageBind to align text， visual， and audio inputs into a shared feature space， obtaining robust multimodal content representations. 3.Interaction Path: Combines LightGCN to capture long-term structural preferences and integrates the sequential model SASRec to absorb short-term contextual preferences， balancing the graph-based diffusion effect and recent sequential dynamics. The embeddings from the three paths are mapped to a latent space of the same dimensionality through independent projection layers. A cross-path contrastive learning approach is then applied for comprehensive alignment (encompassing Semantic-Content， Semantic-Interaction， and Interaction-Content). Finally， a weighted fusion is utilized to calculate the ultimate preference scores. To optimize model training， this study adopts the Bayesian Personalized Ranking (BPR) loss in the interaction path and designs a multi-task contrastive loss (InfoNCE) during the alignment stage to construct a consistent semantic space. Experiments employ the Short-video dataset (WWW 2025) and MicroLens as primary data sources， encompassing short-video content as well as large-scale， multi-typed user interaction behaviors. On the evaluation metrics Hit@K， NDCG@K， and MRR， the proposed method achieves an overall improvement. Compared to existing graph-based recommendation， multimodal modeling， and sequential methods， this framework demonstrates systematic improvements in semantic interpretability， cross-modal consistency， and the unified modeling of long- and short-term preferences. Meanwhile， its design echoes recent research trends， such as the positive effects of neighborhood aggregation representation learning in graph neural networks and contrastive learning on representation space alignment. The main contributions of this paper are: proposing an integrated short-video recommendation architecture featuring a "three-path — contrastive alignment — fusion" paradigm; injecting global graph features into a sequential learning model to balance long- and short-term preferences; and utilizing LLM-driven semantic summaries to bridge the content semantic gap， maintaining high usability and stability without the need for expensive annotations.
