| 摘要: | 隨著電子商務平台商品種類的激增，如何在推薦系統中兼顧推薦準確性、探索性與商業價值成為一項重要挑戰。傳統推薦方法多專注於預測使用者下一步可能購買的商品，強調相關性與個人化，然而這類方法往往導致「過度專業化」問題，推薦結果缺乏多樣性與新奇性，限制了用戶探索潛在興趣的空間。同時，多數現有方法忽略了價格這一關鍵變數，無法有效實現追加銷售策略，限制了平台商業價值的提升。儘管部分研究已開始引入商品價格與新奇性概念，但多數未能系統性整合這些異質訊號，亦缺乏有效機制衡量推薦商品對平台利潤的貢獻。為解決上述問題，本文提出一個名為 USRec 的序列式追加銷售推薦架構，綜合考量商品的相關性、價格與新奇性，以最大化推薦商品的追加銷售價值。
USRec 首先基於用戶的購買序列與商品共現關係構建「序列圖」與「共現圖」，並透過 LightGCN 分別學習商品嵌入。接著，透過對比學習將來自兩張圖的嵌入表示對齊與優化，使模型能同時掌握時序依賴與頻率關聯。優化後的商品表示進一步輸入 Transformer 模型進行序列偏好建模，捕捉長距離興趣演變與語意權重，再融合價格向量與基於時間戳計算的新奇性向量，組成最終的追加銷售分數，進行候選商品排序。最終模型在排序階段考量使用者偏好與探索潛力，並導向高價值商品的推薦。
實驗部分使用 Amazon Review Dataset 中的 Appliances 與 Musical Instruments 兩個類別進行評估，包括排序指標（MRR、NDCG）與追加銷售效益指標（UVD）。結果顯示，USRec 在所有指標上均顯著優於傳統與現代基線模型，展現出在推薦準確度與價格提升效益之間的良好平衡。進一步的消融實驗亦驗證 GNN、對比學習與 Transformer 三大模組皆為不可或缺的關鍵設計，且各自對模型穩定性與效能具正向貢獻。
綜合而言，本文的貢獻在於：一、首次於序列推薦中同時整合相關性、價格與新奇性三向量，提出可衡量追加銷售效益的推薦架構；二、設計雙圖結構與對比學習強化商品嵌入表現，提升模型辨識性與泛化能力；三、透過實證驗證模型能穩定提升排序精準度與平台收益潛力。本研究不僅回應了現有推薦技術在商業導向應用上的不足，也為追加銷售的推薦任務提供了具體可行的解法。;With the rapid expansion of product categories on e-commerce platforms, achieving a balance between recommendation accuracy, exploration, and commercial value has become a critical challenge. Traditional recommendation methods primarily focus on predicting the next item a user is likely to purchase, emphasizing relevance and personalization. However, this often leads to the problem of over-specialization, where recommended items lack diversity and novelty, limiting users’ opportunities to explore new interests. Furthermore, most existing approaches neglect price as a key variable, making it difficult to implement effective up-selling strategies and limiting the platform′s potential for commercial gains. Although some studies have introduced the concepts of price and novelty, they typically lack a systematic integration of these heterogeneous signals and fail to provide mechanisms for measuring the profit contribution of recommended items.
To address these issues, we propose a sequential up-selling recommendation framework named USRec, which simultaneously considers item relevance, price, and novelty to maximize the up-selling value of recommended items. USRec first constructs a sequence graph and a co-occurrence graph based on user purchase histories and item co-occurrence patterns, and then employs LightGCN to learn item embeddings from both graphs. Contrastive learning is used to align and optimize embeddings from the two graphs, enabling the model to capture both temporal dependencies and frequency-based associations. The optimized embeddings are then fed into a Transformer to model user sequence preferences and capture long-range interest dynamics. These are further combined with a normalized price vector and a novelty vector based on interaction timestamps to compute a final up-selling score for candidate item ranking. This design encourages personalized yet commercially valuable recommendations.
We evaluate USRec on two product categories from the Amazon Review Dataset: Appliances and Musical Instruments. Experimental results on ranking metrics (MRR, NDCG) and up-selling performance (UVD) demonstrate that USRec significantly outperforms both traditional and state-of-the-art baselines, achieving a strong balance between recommendation accuracy and value uplift. Further ablation studies confirm the essential contributions of all three core components, namely GNN, contrastive learning, and Transformer, to model stability and performance.
In summary, the main contributions of this work are: (1) the first integration of relevance, price, and novelty into a unified sequential recommendation framework for quantifying up-selling effectiveness; (2) the design of a dual-graph structure and contrastive learning mechanism to enhance item embeddings and improve model generalization; and (3) comprehensive empirical evidence demonstrating stable gains in both ranking precision and commercial value. This study addresses key limitations in existing recommendation systems and offers a practical solution for up-selling-oriented recommendation tasks. |
