基於深度學習之多目標推薦系統研發;A Study on the Multi-Objective Recommender Systems Based on Deep Learning

NCUIR > College of Electrical Engineering & Computer Science > Department of Computer Science and information Engineering > Research Project > Item 987654321/82317

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/82317

Title:	基於深度學習之多目標推薦系統研發;A Study on the Multi-Objective Recommender Systems Based on Deep Learning
Authors:	陳弘軒
Contributors:	國立中央大學資訊工程學系
Keywords:	深度推薦模型;矩陣分解;隱性特徵;深度學習;自動編碼器;推薦系統;Deep recommendation model;matrix factorization;latent factors;deep learning;autoencoder;recommender systems
Date:	2020-01-13
Issue Date:	2020-01-13 14:40:00 (UTC+8)
Publisher:	科技部
Abstract:	本計畫將利用深度學習的技術改善傳統矩陣分解推薦模型的先天限制，並建構多目標的推薦引擎。矩陣分解推薦技術至少存在以下幾個問題： 1. 矩陣分解推薦模型假設每個隱性特徵的權重相同，這不見得符合真實狀況。 2. 矩陣分解推薦模型假設每個隱性特徵彼此獨立，這不見得符合真實狀況。 3. 矩陣分解推薦模型設使用者與物品的隱性特徵向量長度相同，這不見得是最佳的編碼方式。 4. 以內積決定使用者與商品間的關係可能過度簡化這兩者的互動。我們考慮使用更高維度的運算表示兩者間的關係，可能的方式如：深度學習、核方法等。本研究將利用深度學習的架構來建置推薦模型，放寬上述的幾項限制。針對上述第一點，我們已經有初步的想法使每個隱性特徵具備不同的權重。初步驗證顯示：新方法在公開測試資料集上的 RMSE 優於其他矩陣分解模型。矩陣分解除了在推薦系統上，還有許多其他應用，如：預測社群網路的新連結、為機器學習的缺失特徵值進行補值、圖片壓縮等。本研究放寬傳統矩陣分解的限制，因此同樣能應用到這些領域，同時，由於新模型更具一般性，將可能讓這些應用呈現更好的效果。 ;This project plans to leverage on the deep learning technology to build a multi-objective recommendation model, which relaxes the constraints of the matrix factorization-based recom- mendation models. The matrix factorization-based recommendation models at least have the following problems. 1. MF-based models assume that the weight of each latent factor is the same, which may not always be a reasonable assumption. 2. MF-based models assume that each latent factor is independent of other factors, which may not always be a reasonable assumption. 3. MF-based models require the number of latent factors of a user equaling the number of latent factors of an item. However, this may not be a good encoding mechanism. 4. We probably over-simplify the interaction between a user and an item if only applying the inner-product operation. Particularly, we should consider higher-order interactions, such as the kernel method or the deep learning related operations. This research attempts to design the recommendation model based on the deep learning architecture to relax these restrictions. For the first issue, we have already developed a prototype such that each latent factor has distinct weight. Initial experimental results show that the new method performs better than the other MF-based approaches on the test dataset in terms of the root-mean-squared-error (RMSE). In addition to recommender systems, MF-based approaches can be applied to a wide range of applications, such as link prediction, imputation of the missing values, image compression, etc. This study may relax the constraints of the traditional MF-based approaches, so the results can also be applied to the above domains. Since the new techniques are more general than the MF-based approaches, we may get better results on the above applications.
Relation:	財團法人國家實驗研究院科技政策研究與資訊中心
Appears in Collections:	[Department of Computer Science and information Engineering] Research Project

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	252	View/Open

社群 sharing

Loading...