本論文旨要為序列性資料建立一推薦系統。主要目標是根據使用者當前的行為來預測接下來可能的行動。為了解決這樣的問題,前人提出的FPMC模型可同時考慮序列行為及使用者偏好。我們汲取相似於FPMC模型的概念,不僅利用當前的行為,更把離目前較遠的序列行為納入考量,建造一全面性的模型。除此之外,我們衍伸FPMC模型,並加入使用者行為的時間資訊,建造一時間感知的模型。在此模型中,我們提出兩個簡單且有效的方法。第一,我們利用矩陣分解來獲取時間與物品間的潛在關係。第二,我們在BPR優化過程中,利用時間因素改進取負樣技巧。在包含音樂及課程的資料集上所做的實驗結果顯示,我們提出的方法比原先的FPMC模型來得好。
In this thesis, we aim at building a recommender system for sequential data. The goal is to predict a user’s next action based on his or her last basket of actions. In order to solve this task, FPMC is proposed by Rendle to model both sequential behavior and user preference. By utilizing similar concept of FPMC, we attempt to construct a generalized model to predict actions based on not only current actions of users but also their farther previous sequential behavior. In addition, we extend FPMC to incorporate temporal information of behavior. In our model, two simple and effective methods are introduced. First, relationship between time and items is captured by exploiting Matrix Factorization. Second, we improve negative sampling technique by taking time constraint into account for solving BPR optimization. Experimental results on two datasets, including music dataset and course dataset, show that our method outperforms state-of-the-art.