請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88390
標題: | 結合深度自動編碼器與資料填補之推薦系統 The Recommendation System Combining Deep Autoencoder and Data Imputation |
作者: | 歐怡君 Yi-Chun Ou |
指導教授: | 蔡政安 Chen-An Tsai |
關鍵字: | 推薦系統,資料稀疏性,資料填補,深度學習,自動編碼器, Recommendation system,Data sparsity,Data imputation,Deep learning,Autoencoder, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 在資訊蓬勃的時代,人們經常難以有效率的在眾多訊息中獲取所需的資訊,因此,推薦系統在資訊過濾上扮演了重要的角色。本研究的目的是建立一個推薦模型,依照使用者的偏好提供商品推薦。本研究提出的推薦模型是由資料填補和神經網絡模型 (自動編碼器)組成。我們使用了兩份數據集,MovieLens 100K和Cellphone數據集,來驗證和比較我們與其他推薦系統模型的性能。其中,評分數有許多缺失值,我們使用三種填補方法 (填零、SimpleImputer和IterativeImputer)處理缺失值。資料集的特徵包括「僅評分數」或「評分數、使用者偏好和人口資訊」。模型訓練中,本研究嘗試了兩種 (Adagrad、RMSprop)分類器和三個不同的學習率 (0.001, 0.01, 0.1),並以均方誤差 (RMSE)來評估模型的準確率。結果顯示,訓練特徵同時放入評分數與使用者的資訊,並以IterativeImputer方法填補缺失值時,我們提出的模型表現優於其他推薦系統。此外,不同的優化器和學習率在性能上沒有顯著差異。整體而言,相較於先前的模型,本研究之準確率進步了13.5%。最後,以商品的餘弦相似度 (Cosine similary)對使用者做商品推薦。 In the information explosion era, it is difficult for people to effectively search for the required information from the massive content. Therefore, recommendation systems play a crucial role in information filtering. The purpose of this study is to build a recommendation model to provide users with product recommendations according to their preferences. In this paper, we propose a recommendation model composed of an imputation method and a neural network model (using autoencoder). Two benchmark datasets, MovieLens 100K and Cellphone datasets, are used to verify and compare our proposed model with other recommendation systems. Three imputation methods (Fill with 0, SimpleImputer, and IterativeImputer) are implemented here for dealing with the large proportion of missing values in the rating scores. The features of the dataset includes “rating numbers only” or “rating numbers, user preference, and demographic information”. During the model training, experiments were performed with two classifiers (Adagrad and RMSprop) and three learning rates (0.001, 0.01, 0.1). The accuracy of the model was evaluated using root mean square error (RMSE). The results show that our proposed model performs better than other recommendation systems when using both rating numbers and user information as training features, along with the IterativeImputer method for missing values. Moreover, the different optimizers and learning rates have no significant difference in performance. Overall, this study shows that our proposed model can improve the accuracy up to 13.5% compared to other models. In addition, the recommendations for users are made based on the cosine similarity of items. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88390 |
DOI: | 10.6342/NTU202301570 |
全文授權: | 未授權 |
顯示於系所單位: | 統計碩士學位學程 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 目前未授權公開取用 | 8.7 MB | 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。