過去，協同過濾（CF）技術在推薦系統中得到了廣泛的應用。其中由於潛在因子模型擁有將評分值分解為用戶表示向量和項目表示向量的乘積，並用以進行推薦之特性，許多研究者將其應用於深度學習架構下的協同過濾推薦系統之研究。 不同於以往之研究，本研究對於深度學習架構下的潛在因子模型提出了三個方面的改進。首先對於不同的目標項目，由於關注目標的變化，用戶表示向量應該根據目標項目的內容進行動態調整。因此我們在推薦系統中加入了注意力機制，可以根據用戶的歷史購買經驗與目標產品的關係動態調整用戶表示向量。 除此之外，考慮到在現實中人們的偏好通常會隨著時間而改變。因此基於上述注意力模型，本研究應用了兩個時間衰變函數來強調用戶的近期偏好變化。第一個衰變函數考慮的是最近的評分行為比很久以前的評分行為更重要的情況，因為最近的評分行為反映了用戶當前的興趣。第二時間衰變函數考慮了用戶通常更喜歡最近發布的電影而不是很久以前發布的電影的情況。 通過將前述兩個時間衰減函數與注意力模型相結合，我們提出了一種用於項目得分預測的時間衰變自適應潛在分解機（TDADLFM）模型，並將此模型應用於集成 Movielens-10M 和 HetRec2011 的數據集，證明三個因素皆可以提高推薦性能。 ;In the past, collaborative filtering (CF) technology has been widely used in recommendation systems. Usually, the latent factor model is used as the basis for implementing CF recommendation in the deep learning system. It will decompose the rating value into the product of the user embedding vector and the item embedding vector. This study is different from previous studies in three aspects. First, for different target items, due to changes in focus, the user embedding vector should be dynamically adjusted according to the content of the target item. Therefore, we have added an attention mechanism to the recommendation system, which can dynamically adjust the user′s embedding vector according to the relationship between the user′s historical purchase experience and the target product. However, in reality, people′s preferences usually change over time. Therefore, based on the above attention model, this study considers two time decay functions to emphasize the user′s recent preferences. The first decay function considers the situation where the most recent rating is more important than the rating a long time ago, because the most recent rating reflects the user′s current interest. The second time decay function takes into account the situation where users generally prefer recently released movies to movies released a long time ago. By combining these two time decay functions with the attention model, we propose a time decay adaptive latent decomposition machine (TDADLFM) model for item score prediction. This study applies this model to a dataset integrating Movielens-10M and HetRec2011, and proves that all three new considerations can improve recommendation performance.
