在電子商務系統中，推薦一直是一個重要的課題，也因此現今有許多研究專注於如何針對一般序列或session-based序列來推薦客戶下一購買物品，但這些研究卻沒有同時預測下一物品的可能購買時間，少了此時間資訊，將造成經營者不知道應該何時向客戶採取推薦行動的困境。此外，近幾年全局的用戶資訊也廣為被研究使用，因為加入全局資訊可以挖掘出更多可利用的協作訊息，幫助做更有效的推薦。所以本文利用用戶及物品的交互建立全局資訊，改善局部資訊的不足，並運用購買序列資料中的時間資訊，同時預測客戶下一購買物品及購買時間。本文是第一個提出以用戶購買資訊，同時預測購買物品及購買時間，同時結合圖嵌入形成的全局資訊及局部資訊，並利用 Transformer 幫助預測，本研究在真實世界的電子商務資料集進行實驗，大量的實驗表明我們提出的深度網路架構性能優於幾種最先進的推薦方法。;Recommendation has always been an important topic in the field of e-commerce. Therefore, many studies today focus on how to recommend customers′ next purchases for sequential or session-based sequences, but these studies do not also predict when the next item is likely to be purchased. Without this timing information, operators are faced with the dilemma of wondering when they should make a recommendation to their customers. Furthermore, global user information has been extensively studied and used in recent years, as adding global information can explore more collaborative information that can be used to help make more effective recommendations. Therefore, in this study, we exploit user interactions with items to construct global information to improve local information deficiencies, And use the time information in the purchase sequence data to predict the customer′s next purchase and purchase time. This paper first proposes to use user purchase information, combined with global information and local information formed by graph embedding, and then use transformer to help predict, and finally get the predicted purchase items and purchase time. This study conducts experiments on a real-world e-commerce dataset. Extensive experiments show that our proposed deep network architecture outperforms state-of-the-art recommendation methods.
