研究基於知識蒸餾與交叉注意力機制之跨領域推薦系統;A Study of Cross-Domain Recommendation Based on Knowledge Distillation and Cross-Attention

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/95401

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/95401

Title:	研究基於知識蒸餾與交叉注意力機制之跨領域推薦系統;A Study of Cross-Domain Recommendation Based on Knowledge Distillation and Cross-Attention
Authors:	林方瑜;Lin, Fang-Yu
Contributors:	資訊管理學系
Keywords:	跨領域推薦系統;交叉注意力機制;知識蒸餾;資料稀疏;冷啟動問題;Cross-domain Recommendation System;Cross-attention;Knowledge Distillation;Data Sparsity;Cold Start Issue
Date:	2024-03-14
Issue Date:	2024-10-09 16:46:34 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著電子商務的發展，線上購物已經成為現代生活中不可或缺的一部份，而推薦系統在這之中扮演著相當重要的角色，然而傳統的推薦系統在資料量較少的領域，容易面臨到「資料稀疏(Data Sparsity)」與「冷啟動(Cold Start)」問題，跨領域推薦系統能夠很好的改善這個問題，本研究提出兩個跨領域推薦模型，皆使用自注意機制(Self-attention) 來動態捕捉使用者在不同領域的偏好，並分別使用交叉注意力機制(Cross-attention)與知識蒸餾(Knowledge Distillation)來將知識從來源領域移轉到目標領域，以改善傳統推薦系統在目標領域上遇到問題，提升目標領域的推薦效果，最後使用來自現實世界的 Amazon 數據集進行評估，實驗結果顯示與其他跨領域推薦系統相比，我們研究中提出的兩個模型，在三個指標與兩組資料集上平均提升了 8.49%與 4.78%。;Recent years have seen a burgeoning of recommender systems and application of that recommender systems to E-commerce. Online shopping has become an indispensable part of modern life. The recommender system plays a crucial role in E-commerce. However, the main difficulty surrounding traditional recommender systems is the data sparsity and cold start issues in the domain with fewer data. In this study, we propose two different cross-domain recommendations, both of which utilize Self-attention to dynamically capture user preferences in different domains and implement Cross-attention and Knowledge Distillation to transfer knowledge from the source domain to the target domain respectively. The major purpose of our models is to improve the problem encountered by the traditional recommendation and enhance the effect of recommendation in the target domain. Last but not least, we utilize the real-world Amazon dataset to evaluate our two models. The experimental results show a striking improvement of our models on two datasets and three evaluation matrices in comparison to other state-of-the-art recommendation models.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	19	View/Open

社群 sharing

Loading...