基于张量相似度的推荐方法研究

Research on Recommendation Method Based on Tensor Similarity

传统推荐方法中将用户建模为向量的建模方式只关注用户单方面偏好,为了弥补此种建模方法的局限性,提出一种将用户建模为矩形的张量建模方法。构建了一个基于融合协同过滤与序列推荐算法的推荐模型,该模型集成了Fastformer模型和键值记忆网络对用户张量进行建模;结合用户张量与目标物品的距离及偏置项对用户张量与目标物品的相似度进行计算。在MovieLens和CiaoDVD数据集上对该模型进行实验验证,实验结果表明,该模型能够关注用户多方面偏好并在推荐结果的精准度上优于基线方法,特别是在HR与NDCG评价指标上分别比现有基线方法平均提高了1.4%、1.95%。

Traditional recommendation methods model users as vectors in a way that focuses only on unilateral user preferences.In order to compensate for the limitations of this modelling approach,a tensor modelling method is proposed that models the user as a rectangle.Firstly,a recommendation model based on a fusion of collaborative filtering and sequential recommendation algorithms is constructed,which integrates the Fastformer model and key-value memory network to model the user tensor;secondly,the similarity between the user tensor and the target item is calculated by combining the distance between the user tensor and the target item and the bias term.The model is experimentally validated on the MovieLens and CiaoDVD datasets,and the results show that the model is able to focus on multiple user preferences and outperforms the baseline method in terms of accuracy of recommendation results.In particular,the HR and NDCG evaluation metrics are improved on average by 1.4%and 1.95%,respectively,over the existing baseline method.