| 摘 要: 针对显式和隐式反馈不足而导致个性化推荐偏差的问题,提出一种基于移动平均值反馈美学的对抗性协同过滤推荐算法(ACF-MA)。首先,通过整合正反馈、负反馈及未观测样本的三元组数据,构建多维度商品评分框架以精准建模用户偏好;其次,增加美学因素模型强化对商品偏好的特征提取,并构建对抗性学习方法优化减轻反馈噪声的影响;最后,在亚马逊的Jewelry数据集上展开多组对照实验。结果表明相较于次优算法,在推荐列表长度为50时,所提算法———ACF-MA 在命中率(HR)和归一化折损累计增益(NDCG)评估指标上分别最高提高了15.1%和17.4%,验证了该算法在提升推荐系统准确性方面的有效性。 |
| 关键词: 推荐系统 对抗性学习 视觉感知 个性化排序 |
|
中图分类号: TP391.3
文献标识码: A
|
| 基金项目: 浙江省自然科学基金重点项目资助(LZ22F010005) |
|
| Adversarial Collaborative Filtering Recommendation based on Feedback Aesthetics |
|
ZHU Jianjun, CHEN Boyu, WU Zhefu
|
(College of Information Engineering, Zhejiang University of Technology, Hangzhou 310000, China)
zjj@zjut.edu.cn; 221122030345@zjut.edu.cn; wzf@zjut.edu.cn
|
| Abstract: To address the issue of personalized recommendation bias caused by insufficient explicit and implicit feedback, an Adversarial Collaborative Filtering recommendation algorithm based on feedback aesthetics-Moving Average(MA)(ACF-MA)is proposed. Firstly, a mult-i dimensional product rating framework is constructed by integrating triplet data consisting of positive feedback, negative feedback, and unobserved samples to accurately model user preferences. Secondly, an aesthetic-aware-module is added to strengthen the feature extraction of product preferences, and an adversarial learning method is constructed to optimize and reduce the impact of feedback noise. Finally, multiple comparative experiments are conducted on the Amazon Jewelry dataset. The results show that compared with the state-o-f the-art(SOTA), the proposed algorithm has the highest improvement of 15.1% and 17.4% in the Hit Rate(HR) and Normalized Discounted Cumulative Gain(NDCG) evaluation indicators respectively, verifying the effectiveness of the
algorithm in improving the accuracy of the recommendation system. |
| Keywords: recommendation system adversarial learning visual perception personalized ranking |
