普洪飞, 邵剑飞, 张小为, 魏榕剑. 融合动态兴趣偏好与特征信息的序列推荐[J]. 云南大学学报(自然科学版), 2022, 44(4): 708-717. doi: 10.7540/j.ynu.20210542
引用本文: 普洪飞, 邵剑飞, 张小为, 魏榕剑. 融合动态兴趣偏好与特征信息的序列推荐[J]. 云南大学学报(自然科学版), 2022, 44(4): 708-717. doi: 10.7540/j.ynu.20210542
PU Hong-fei, SHAO Jian-fei, ZHANG Xiao-wei, WEI Rong-jian. Sequence recommendation of fusing dynamic interest preference and feature information[J]. Journal of Yunnan University: Natural Sciences Edition, 2022, 44(4): 708-717. DOI: 10.7540/j.ynu.20210542
Citation: PU Hong-fei, SHAO Jian-fei, ZHANG Xiao-wei, WEI Rong-jian. Sequence recommendation of fusing dynamic interest preference and feature information[J]. Journal of Yunnan University: Natural Sciences Edition, 2022, 44(4): 708-717. DOI: 10.7540/j.ynu.20210542

融合动态兴趣偏好与特征信息的序列推荐

Sequence recommendation of fusing dynamic interest preference and feature information

  • 摘要: 传统的序列推荐通常忽略用户和项目特征信息的重要性,且无法有效对动态的兴趣偏好进行建模. 因此,提出融合动态兴趣与特征信息的序列推荐算法. 该算法通过对目标项目进行动态兴趣建模,克服兴趣转移带来的影响;同时融合用户和项目特征信息模拟真实的用户行为以提高推荐的性能. 首先,针对动态兴趣建模,采用辅助函数应用下一个行为监督上一个隐藏兴趣状态的学习,并采用带注意力机制的门控循环单元为不同的兴趣状态对目标影响程度赋予不同的权重;然后,针对用户和项目特征信息特征融合,采用平凡注意力机制为影响目标项目的特征赋予不同的权重,并通过多头注意力机制进行深层次的特征提取;最后,融合用户动态兴趣表示和用户项目特征表示输入到多层感知机. 在Yelp和 MovieLens-1M数据集上进行仿真实验,结果表明提出模型的性能比一些基线模型有较好的提升.

     

    Abstract: Traditional sequence recommendation usually ignores the importance of user and item feature information, and cannot effectively model dynamic interest preferences. Therefore, this paper proposes a sequence recommendation that integrates dynamic interest and feature information. The algorithm overcomes the impact of interest transfer by modeling the target item's dynamic interest; at the same time, it integrates user and item feature information to simulate real user behavior to improve recommended performance. Firstly, for dynamic interest modeling, an auxiliary function is used to apply the next behavior to supervise the learning of the previous hidden interest state, and a gated cycle unit with attention mechanism is used to assign different weights to the degree of influence of different interest states on the target. Then, aiming at the fusion of user and item feature information, the vanilla attention mechanism is used to assign different weights to the features that affect the target item, and the multi-head attention mechanism is used for in-depth feature extraction. Finally, the user's dynamic interest representation and user item feature representation are input to the multi-layer perceptron. Simulation experiments are carried out on the datasets of Yelp and MovieLens-1M. The results show that the performance of the proposed model outperforms some baseline models.

     

/

返回文章
返回