基于动态卷积概率矩阵分解的潜在群组推荐

发布时间：2018-07-22 17:14

【摘要】：近年来,群组推荐由于其良好的实用价值得到了广泛关注.然而,已有的群组推荐方法大多都是根据分析用户对服务的评分矩阵直接将个体用户的推荐结果或个体用户偏好进行聚合,没有综合地考虑用户-群组-服务这三者间的联系,导致群组推荐效果欠佳.受潜在因子模型与状态空间模型启发,结合评分矩阵、服务描述文档以及时间因素,共同分析用户-群组-服务间的联系,提出了一种基于动态卷积概率矩阵分解的群组推荐方法.该方法首先利用基于卷积神经网络的文本表示方法获取服务潜在特征模型的先验分布;然后,将状态空间模型与概率矩阵分解模型相结合,获得用户潜在偏好向量与服务特征向量;之后,对用户偏好向量运用聚类算法来发现潜在的群组;最终,对群组中的用户偏好采取均值策略融合成群组偏好向量,并与服务特征向量共同生成群组对服务的评分,实现群组推荐.通过在MovieLens数据集上与同类方法进行对比实验,发现所提方法的推荐有效性与精确性上更具有优势.
[Abstract]:In recent years, group recommendation has received wide attention because of its good practical value. However, most of the existing group recommendation methods aggregate the individual users' recommendation results or individual user preferences directly based on the analysis of users' rating matrix, and do not consider the relationship between user, group and service. As a result, the group recommendation effect is not good. Inspired by the latent factor model and the state space model, the relationship between users, groups and services is analyzed by combining the scoring matrix, service description documents and time factors. A group recommendation method based on dynamic convolution probability matrix decomposition is proposed. Firstly, the prior distribution of service potential feature model is obtained by the text representation method based on convolution neural network, and then the state space model is combined with the probability matrix decomposition model. The user potential preference vector and the service feature vector are obtained. After that, the clustering algorithm is applied to the user preference vector to find the potential group. Finally, the mean value strategy is used to fuse the user preferences in the group into the group preference vector. The group recommendation is realized by generating the service score together with the service feature vector. By comparing with the similar methods on the MovieLens dataset, it is found that the proposed method is more effective and accurate.
【作者单位】：南京邮电大学计算机学院;江苏省无线传感网高技术研究重点实验室;江苏省大数据安全与智能处理重点实验室;
【基金】：国家自然科学基金项目(61201163,61373138)~~
【分类号】：TP391.3

【参考文献】