Robust multi-objective visual bayesian personalized ranking for multimedia recommendation

Machine learning classifiers are susceptible to adversarial perturbations, and their existence raises security concerns with a focus on recommendation systems. While there is a substantial effort to investigate attacks and defensive techniques in recommendation systems, Basic Iterative perturbation...

Full description

Saved in:
Bibliographic Details
Published in:Applied intelligence (Dordrecht, Netherlands) Netherlands), 2022-03, Vol.52 (4), p.3499-3510
Main Authors: Paul, Agyemang, Wu, Zhefu, Liu, Kai, Gong, Shufeng
Format: Article
Language:English
Subjects:
Aesthetics
Artificial Intelligence
Computer Science
Customization
Iterative methods
Machine learning
Machines
Manufacturing
Mechanical Engineering
Minimax technique
Multimedia
Multiple objective analysis
Perturbation
Processes
Ranking
Recommender systems
Training
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Machine learning classifiers are susceptible to adversarial perturbations, and their existence raises security concerns with a focus on recommendation systems. While there is a substantial effort to investigate attacks and defensive techniques in recommendation systems, Basic Iterative perturbation strategies (BIM) have been under-researched in multimedia recommendation. In this work, we adapt the iterative approach for multimedia recommendation. We proposed a novel Dynamic Collaborative Filtering with Aesthetic (DCFA) approach which leverages aesthetic features of clothing images into a multi-objective pairwise ranking to capture consumer aesthetic taste at a specific time through adversarial training (ADCFA). The DCFA method extends visual recommendation to make three key contributions: (1) incorporate aesthetic features into multimedia recommender system to model consumers’ preferences in the aesthetic aspect. (2) Design a multi-objective personalized ranking for the visual recommendation. (3) Use the aesthetic features to optimize the learning strategy to capture the temporal dynamics of image aesthetic preferences. To reduce the impact of perturbation, we train a DCFA objective function using minimax adversarial training. Extensive experiments on three datasets demonstrate the effectiveness of our method.
ISSN:0924-669X
1573-7497
DOI:10.1007/s10489-021-02355-w