Novel Distributed Multimedia Recommendation Systems Using Personalized Information

In this paper, we propose a novel distributed multimedia recommendation system (DMRS) to address personalized preference by use of the matrix-sketching approach for dimensionality reduction and local information updates. Conventional recommendation systems can hardly address scalability, privacy, an...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on broadcasting pp. 1 - 12
Main Authors: Chang, Shih Yu, Wu, Hsiao-Chun, Yan, Kun, Huang, Scott Chih-Hao, Wu, Yiyan
Format: Journal Article
Language:English
Published: IEEE 2025
Subjects:
Collaborative filtering
Digital multimedia broadcasting
Dimensionality reduction
Distributed databases
Distributed multimedia recommendation system (DMRS)
Electronic mail
matrix sketching
Multimedia systems
normalized root-mean-square error (NRMSE)
Privacy
randomized matrix-factorization algorithm
Recommender systems
Robustness
Servers
ϵ-robustness
ISSN:0018-9316, 1557-9611
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In this paper, we propose a novel distributed multimedia recommendation system (DMRS) to address personalized preference by use of the matrix-sketching approach for dimensionality reduction and local information updates. Conventional recommendation systems can hardly address scalability, privacy, and robustness, all of which are very important in practice. To combat the aforementioned challenges, we propose to incorporate local information updates based on local private information at the client side to protect privacy by restricting users' data from the server and utilize the matrix-sketching scheme to further reduce the dimensionality of the global user-item interaction data so that the personalized (distributed) recommendations can be made by users' devices in local. To evaluate the system performance, we define a new robustness measure, namely ϵ-robustness, which quantifies the performance consistency of the recommendation system and involves both sketching errors and local rating updates. Furthermore, we introduce a novel randomized matrix-factorization algorithm to achieve the desired robustness while still maintaining the interaction-data fidelity in terms of normalized root-mean-square error (NRMSE). Our experimental results on both simulated and real-world data demonstrate the effectiveness of our proposed novel DMRS in attaining a good balance between the interaction-data fidelity and the system robustness subject to the privacy protection.
ISSN:0018-9316
1557-9611
DOI:10.1109/TBC.2025.3583989