請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/77811
標題: | 抽樣策略對成對隱含狄利克雷分布推薦系統成效影響之探討 The Effect of Quality Pair Selection to Pairwise Latent Dirichlet Allocation Recommendations |
作者: | Fang-Yu Lin 林芳瑀 |
指導教授: | 陳建錦 |
關鍵字: | 推薦系統,偏好學習,成對學習,排序學習,隱含狄利克雷分布, Recommendation Systems,Learning to Rank,Preference Learning,Pairwise learning,Latent Dirichlet Allocation, |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 推薦系統的主要功能為自動提供對目標消費者有用的產品資訊。系統透過分析消費者的購買歷史紀錄、消費者的特性或產品的特性來了解消費者的行為或偏好。根據推薦系統自動提供的資訊,消費者降低了搜尋的成本,並增加了消費者購買他們感興趣的產品的可能性,這顯示了持續改進推薦系統的重要性。我們使用隱含狄利克雷分布演算法來獲取消費者及商品之偏好,並將其與成對學習的概念相結合。我們將任兩個消費者評分過的商品組合成配對,而商品在配對中的排序,來自於消費者給予商品的評分。演算法從這些配對中學習消費者對商品的偏好,並藉此預測未被評分過的商品的排序。實驗結果顯示,改進後的方法能將使用者可能會喜歡的商品排序在推薦清單中較前面的位置。此外,為了在計算上實現更好的效率,我們還引入了幾種抽樣策略,不僅減少了數據集的大小,還避免了使用資訊量較低的配對,改進的推薦模型得以更精準的推薦使用者可能有興趣的商品。 A recommendation system is developed to automatically provide product information that could be useful to the target consumer. The system learns the behavior or the preference of the consumers by their purchase history, their profile or the characteristics of the product. According to the information the recommendation system automatically provides, the consumers lower the cost of searching and increase the possibility of consuming the products they are interested in, which demonstrates the importance of continuous refinement of the recommendation system. We use latent Dirichlet allocation (LDA) (Blei et al., 2003) to acquire user and item preferences and combine it with the concept of pairwise learning by defining the precedence of items in a pair as an item is preferred than another if its rating is higher than the other’s rating. The experiment results show that our improved method achieves better performance in terms of the recommendation precision and MRR than many well-known recommendation methods. In order to achieve better efficiency, we also introduce several sampling strategies which not only decrease the size of dataset but also avoid using least informative data to build a better recommendation model. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/77811 |
DOI: | 10.6342/NTU201801054 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊管理學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-107-R05725009-1.pdf 目前未授權公開取用 | 788.96 kB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。