因為電子商務的逐漸普及,推薦系統成為使用者尋找所需商品的重要角色。推薦系統主要是預測並推薦下一項商品給使用者,並且準確率愈高愈好,然而因為使用者喜好一般來說都不是固定而是隨著時間變換的,所以如何依據使用者的過去行為去對使用者的動態改變的喜好進行建模就成為一項重要性和挑戰性兼具的課題。以前的方法大部分使用序列神經網路,將使用者從左到右的歷史交互關係編碼以進行推薦,然而這樣單向的架構會限制住使用者行為序列中隱藏表示的能力,而且在現實生活中一個人的行為並非井然有序的序列,因此雙向模型被提出,其中,採用完形填空任務有效地避免訊息洩漏。另外,我們認為商品的類別以及使用者類型的相似度越高,可能會喜好同一項商品的機率也會越高,因此如果我們能善用這些資訊勢必能增加推薦系統的準確率。所以本論文計畫透過雙向模型架構融合前後訊息,並加入了產品特徵及使用者的相關資訊,使得推薦系統能夠更加精準。透過在一些真實的數據集上進行的實驗結果也顯示,我們所提出的模型架構優於目前幾種較常被使用的推薦方法,證明了其推薦系統模型的實用性。;The popularity of e-commerce makes recommendation systems a necessary tool for users to find the commodities of desire. For recommender systems, modeling the user’s dynamic preferences based on historical behavior is important, but also challenging. The previous methods used sequential neural networks to encode the user’s left-to-right historical interaction relationship for the recommendation. However, the unidirectional architecture has limited capability to hide representations in user behavior sequences and rigidly ordered sequences were not realistic. Therefore, a bidirectional model is proposed, and the cloze is employed to efficiently train the model to avoid information leakage. In this paper, adopting the bidirectional model to integrate the left-and-right information, and add relevant information about item characteristics and users (BTSR), makes more accurate recommendations. We have conducted several experiments on four real datasets, empirical evidence shows that BTSR′s recommendation outperforms other state-of-the-art baseline models. Furthermore, demonstrate the practicability of the proposed model recommendation system.