| 摘 要: 为了改进时间事件上的输入扰动或者攻击可能导致系统推荐性能大幅下降的问题，提出一种基于对抗性训练改进模型鲁棒性的协同过滤推荐算法。通过构建微小扰动对推荐模型进行训练，调整改进网络结构参数，从而提高系统的推荐准确度和抗干扰能力。通过在亚马逊数据集上的实验，并与几个基线模型进行不同Top-K推荐目标下的NDCG性能对比，结果表明：经过对抗训练的改进算法提升了系统鲁棒性，并且在中等扰动情况下可减少性能下降15%以上。 |
| 关键词: 协同过滤推荐；鲁棒性；对抗性训练 |
|
中图分类号: TP182
文献标识码: A
|
| 基金项目: 浙江省自然科学基金项目(LY18F010025,LY13F010011). |
|
| Dynamic Collaborative Filtering Recommendation Algorithm based on Adversarial Training |
|
HUANG Daqiao1, ZHU Jianjun2, CAO Junzuo2
|
( 1.Zhejiang Communication Service Network Technology Branch, Hangzhou 310000, China ;
2.College of Information Engineering, Zhejiang University of Technology, Hangzhou 310000, China)
huangdaqiao@126.com; zjj@zjut.edu.cn; 2257416330@qq.com
|
| Abstract: Aiming at the problem that input disturbance or attacks on time events may lead to a significant drop in system recommendation performance, this paper proposes a collaborative filtering recommendation algorithm based on adversarial training to improve model robustness. The recommendation model is trained by constructing small disturbances, and the network structure parameters are adjusted and improved, thereby improving the recommendation accuracy and antiinterference ability of the system. Through experiments on the Amazon dataset and comparison of its NDCG (Normalized Discounted Cumulative Gain) performance with that of several baseline models under different Top-K recommendation targets, the results show that the improved algorithm after adversarial training enhances the robustness of the system and reduces the performance degradation by more than 15% under moderate disturbance. |
| Keywords: collaborative filtering recommendation; robustness; adversarial training |
