近年來，次購物籃推薦（NBR）因具有廣泛的應用價值而受到高度關注。相較於以單一商品為單位的序列推薦，次購物籃推薦能夠建模同一筆交易中共同購買的商品，並更完整的捕捉使用者的購買意圖。然而，現有方法仍面臨多項挑戰，包括稀疏的購買歷史、不規律的時間間隔、重複消費與新商品探索等問題。本研究提出一個結合反向增強與時間對比學習的次購物籃推薦框架 RABRec。該框架透過反向資料增強擴充較短的購買歷史，並利用考量購物籃間時間間隔的對比學習目標緩解冷啟動問題。此外，RABRec 將持續性購買意圖與偏好漂移加以區分，並設計情境調節式重複購買殘差機制。此機制能根據使用者的購買情境，自動校正曾購買商品的預測分數，同時維持新商品的競爭力。在多個真實世界資料集上的實驗結果顯示，RABRec 在多數評估指標上均優於現有方法。最後，本研究透過多項案例分析，進一步驗證所提出 RABRec 框架的實務可行性。;Recently， Next-Basket Recommendation (NBR) has attracted tremendous attention due to its widespread applicability. NBR models items purchased together within a transaction and captures purchase intentions better than item-level sequential recommendation. However， existing methods still struggle with several challenges， e.g.， sparse histories， irregular time intervals， repeat consumption， new-item exploration， to name a few. In this study， we propose a reverse-augmented temporal contrastive framework， RABRec， which augments short histories and alleviates cold-start issues via time-aware contrastive objective over inter-basket time intervals. In addition， RABRec separates continuous intent from preference drift and introduces a context-gated repeat residual mechanism to adaptively calibrate scores of previously purchased items while keeping new items competitive. Experiments on various real-world datasets indicate that RABRec outperforms state-of-the-art methods on most evaluation metrics. Finally， we give several case studies to show the practicability of the proposed RABRec framework.
