一种融合用户动态偏好和注意力机制的跨领域推荐方法

Approach to Cross-domain Recommendation Fusing Users′Dynamic Preferences and Attention Mechanism

作为当今电子商务中的一项重要技术,推荐系统的重要性日益提升.在项目空间上用户的评分数据十分稀疏,导致推荐系统的质量不佳.商品评论中蕴含着丰富的信息,通过提取评论文本信息能够有效地减少数据稀疏性带来的影响.事实上,用户的偏好并非一成不变的,将不同时间段设置不同的权重能更有效地描述用户的整体状况.在神经网络算法广泛应用的背景下,将神经网络引入到跨领域推荐中可以发现不同领域用户偏好的映射关系.此外,注意力机制是一种流行的深度学习方法,将注意力机制与主题模型结合,提出一种基于注意力机制的跨领域推荐方法.首先,使用LDA(Latent Dirichlet Allocation)主题模型分别提取源领域和目标领域的项目主题分布.接着,将其与用户评分、时间权重因子、注意力机制结合,得到用户的动态偏好.然后,使用BP(Back Propagation)神经网络学习用户偏好的映射关系,并将用户在源领域与目标领域的偏好结合.最后,通过协同过滤的方法进行评分预测.实验结果表明,提出的推荐方法在亚马逊电子商品、影视与以及音乐的评分评论数据集上较其它传统推荐策略有着更好的推荐效果.

As an important technology in today′s e-commerce,recommendation system is becoming more and more important.The rating data of users is very sparse in project space,which leads to poor quality of recommendation system.Commodity reviews contain abundant information,and the impact of data sparseness can be effectively reduced by extracting review text information.In fact,users′preferences are not static,and setting different weights in different time periods can describe the overall situation of users more effectively.With the wide application of neural network algorithm,the mapping relationship of user preferences in different fields can be found by introducing neural network into cross-domain recommendation.In addition,attention mechanism is a popular deep learning method.Combining attention mechanism with topic model,it proposes a cross-domain recommendation method based on attention mechanism.Firstly,LDA(Latent Dirichlet Allocation)theme model is used to extract the project theme distribution of source domain and target domain respectively.Then,it is combined with user score,time weighting factor and attention mechanism to get the dynamic preference of users.Then,BP(Back Propagation)neural network is employed to learn the mapping relationship of user preferences,and the user preferences in the source domain and the target domain are combined.Finally,the score is predicted by collaborative filtering.The experimental results show that the proposed recommendation method has better recommendation effect than other traditional recommendation strategies on Amazon′s electronic products,movies and music rating and comment data sets.