| 摘 要: 在推荐系统中,用户行为数据通常存在高度稀疏性,并在对抗扰动下易导致性能下降,从而削弱结果可靠性。为此,提出一种信息增强的多样本对比学习方法,通过多视图表征与对比损失提升短序列行为的表征能力。在此基础上,进一步构建融合多任务学习与对抗训练的鲁棒推荐模型,引入自注意力机制提取扰动增强特征,并结合对抗对比损失与自适应损失权重提升抗扰性。实验结果表明,该模型在 HR和 NDCG指标上较基线方法平均提升1.60%~3.13%,并在ε=0.3扰动下将 NDCG@10的下降幅度由46.4%降至18.2%,显著提升推荐的准确性与鲁棒性。 |
| 关键词: 数据稀疏 对抗扰动 可信推荐 多任务学习 对抗对比学习 |
|
中图分类号: TP3-0
文献标识码: A
|
|
| Research on Trustworthy Recommendation Model Based on Information-Enhanced Contrastiv eLearning |
|
ZHU Yue1, YIN Pei1,2, ZHAO Jun1, QUAN Guanting1
|
(1.Business School, University of Shanghai for Science & Technology, Shanghai 200093, China;
2.School of Intelligent Emergency Management, University of Shanghai for
Science and Technology, Shanghai 200093, China)
zhuyue1118@126.com; pyin@usst.edu.cn; 212421080@st.usst.edu.cn; 2335061617@st.usst.edu.cn
|
| Abstract: In recommendation systems, user behavioral data is typically highly sparse and prone to performance degradation under adversarial perturbations, thereby undermining the reliability of results. To address this, this paper proposes an information-enhanced mult-i sample contrastive learning approach, which improves the representational capacity of short behavioral sequences through mult-i view representations and contrastive losses. Building on this foundation, a robust recommendation model integrating mult-i task learning and adversarial training is developed. Asel-f attention mechanism is incorporated to extract perturbation-enhanced features, while adversarial contrastive losses and adaptive loss weighting are employed to strengthen resilience against disturbances. Experimental results demonstrate that the proposed model achieves average improvements of 1.60% to 3.13% over baseline methods on HR and NDCG metrics, and reduces the NDCG@10 drop underε=0.3perturbations from 46.4% to 18.2% , significantly enhancing both recommendation accuracy and robustness. |
| Keywords: data sparsity adversarial perturbation trustworthy recommendation mult-i task learning adversarial
contrastive learning |
