現今最流行處理大量數據集的方法之一是矩陣分解(MF)技術，矩陣分解常用於在推薦系統中，因為其預測用戶興趣有著非常高的準確度。特別是非負值矩陣分解(NMF)已經被證明能夠非常有效利用多變量數據集的分解。然而即使是NMF技術，也無法完全捕捉到時間對於用戶喜好的影響程度。 本研究利用使用者因為時間影響而對喜好的改變，我們提出兩個基於傳統NMF的創新推薦系統 Dec_NMF，透過有效的時間影響，考慮使用者對於喜好的改變。Dec_NMF 包含了人類喜好行為隨著時間改變的概念，考慮使用者目前喜好的偏好，並將評分時間過長的資訊做衰減的處理。 本研究立用了三種線性以及三種非線性函數來調整評分，設定不同的潛在因素數量，透過均方誤差、均方根誤差、平均絕對誤差指標來評估模型好壞，並使用準確度、精密度、召回率與F1方法衡量模型的效能。實驗結果表明提出的方法中，在大部分的狀況優於傳統非負值矩陣分解。此外我們將所提出的模型應用在MovieLens資料集上，來演示所提出模型的有效性。 ;One of the most popular approaches to handle very large datasets is matrix factorization(MF) technique. The MF method was commonly used in recommendation systems due to the precise prediction of the user’s interest. Especially one of the successful method Non-Negative Matrix Factorization (NMF) has previously been shown to be a useful decomposition for multivariate data. NMF-based techniques, however, could not properly capture time influences on user preferences. In this paper, by considering time impacts on preferences, we propose two novel NMF-based recommendation system, Dec_NMF, to consider user preferences over time. Our proposed method extends the concept of the change of human interest through time to capture user’s current preference and reduce impacts which was rated from a long time ago. We adjust the rating using three different linear and three different non-linear time decay. Each function represents different decay degree of preferences to simulate the human’s interest behavior. The experimental results show that proposed methods outperforms the traditional MF and NMF model. Furthermore, we apply Dec_NMF on MovieLens datasets to demonstrate the effectiveness of Dec_NMF recommendation.
