计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (9): 272-279.DOI: 10.3778/j.issn.1002-8331.2205-0437

• 大数据与云计算 • 上一篇    下一篇

融合双层注意力机制的群组偏好融合策略研究

梅雨竹,胡竹林,朱欣娟   

  1. 1.西安工程大学 计算机科学学院,西安 710000
    2.陕西省图书馆数字资源部,西安 710000
  • 出版日期:2023-05-01 发布日期:2023-05-01

Research on Group Preference Fusion Strategy Based on Two-Layer Attention Mechanism

MEI Yuzhu, HU Zhulin, ZHU Xinjuan   

  1. 1.School of Computer Science, Xi’an  Polytechnic University, Xi’an 710000, China
    2.Digital Resources Department of Shaanxi Provincial Library, Xi’an 710000, China
  • Online:2023-05-01 Published:2023-05-01

摘要: 当前推荐系统研究热点及其演变趋势之一是个性化推荐由关注个体推荐逐步转向关注群体推荐。目前多数群组推荐方法在选择偏好融合策略时习惯采用预定义的静态策略,而静态策略的特点就导致算法无法最大化模拟出群组决策的真实过程。在前人研究的基础之上提出一种基于双层注意力机制的群组推荐方法,该方法充分考虑到群体用户的差异性和相互影响,以及对于不同领域的决策权等问题。计算群组内每位成员对其他成员的注意力权重,获得群组成员特征向量,再计算每个成员在选择某一个项目的注意力权重,为群组生成对于该项目的偏好向量,以此来充分还原群组用户之间的交互以及群组决策的过程。通过在CAMRa2011和Meetup数据集上与COM、SIG、AGR、AGREE、FastGR等方法在不同参数条件下进行了对比,在归一化折扣累计增益和命中率两个指标上,相较基线模型平均提高了0.025 4和0.030 7。

关键词: 群组推荐, 注意力机制, 偏好融合策略, 推荐算法

Abstract: One of the current research hotspots of recommender systems and its evolution trend is that personalized recommendation gradually shifts from focusing on individual recommendation to focusing on group recommendation. At present, most group recommendation methods are accustomed to adopting a predefined static strategy when choosing a preference fusion strategy, and the characteristics of the static strategy make the algorithm unable to maximize the simulation of the real process of group decision-making. On the basis of previous research, this paper proposes a group recommendation method based on a two-layer attention mechanism, which fully takes into account the differences and mutual influence of group users, as well as the decision-making power in different fields. The attention weight of each member in the group is calculated to other members, the group member feature vector is obtained, and then the attention weight of each member in selecting a certain item is calculated, and the preference vector for the item for the group is generated. The interaction between group users and the process of group decision-making are fully restored. By comparing the CAMRa2011 and Meetup datasets with COM, SIG, AGR, AGREE, FastGR and other methods under different parameter conditions, the two indicators of normalized discount cumulative gain and hit rate are higher than the baseline model, is up to 0.025 4 and 0.030 7.

Key words: group recommendation, attention mechanism, preference fusion strategy, recommendation algorithm