Bundle Recommendation Methods Considering Rating Data Differences For Online Supermarkets

Online supermarkets have a large number of users and items.Accurate and effective personalized bundle recommendation can extend the range of users’ selection,increase the platform sales and profitability,and increase user’s satisfaction at the same time.This has great significance to the development of online supermarkets.The rating data could directly reflect the users’ preferred items,and the rating differences could reflect the level of preference.Meanwhile,how to use the differences in rating data and combine with other data to improve the accuracy of bundle recommendation still has some challenges.(1)Data sparsity problem.In e-commerce platforms,compared to the number of users,the number of interactions between those is very limited.(2)Data heterogeneity problem.The structure of user data and item data are different,and there is a data heterogeneity problem in recommendation when all the data are used.(3)The problem of difficulty in mining the relationships between items.The items attribute data has less information describing the correlation between items,which is difficult to mine by using the item attribute data only.In response to the above problems,this thesis proposes a bundle recommendation method considering rating data differences for online supermarkets.The main work is as follows:(1)For the problem of data sparsity and data heterogeneity,this thesis proposes a new hybrid rating prediction method.It considers the differences in rating values and the influence of the number of common rates when calculating user similarity.A rate correction coefficient,a product popularity correction coefficient and a user satisfaction coefficient are proposed to improve the cosine similarity measurement method.Meanwhile,a product graph network transformation data structure is constructed with self-attention mechanism is used to integrate item data to calculate item characteristics.It solves the problem of data sparsity and data heterogeneity to a certain extent.(2)For the problem of difficult to mine the relationships between items,this thesis mines the interaction information of items by using the category attribute,price attribute and interaction information attribute in the rating data and item attributes.The graph self-attention mechanism is used to construct the item graph network and learn the influence weight of the interaction information attribute between items on the relationships which improves the quality of the bundles generated.(3)Numerical experiments and analysis are conducted.The typical e-commerce public dataset is used as an example for experimentation.Through comparative experiments,the effectiveness of the proposed model in the combined recommendation task are verified.In order to prove the influence of parameters on model performance,ablation experiments were performed on the correction factors and models considering low scores.The sensitivity experiments were performed on the parameters such as potential factors,weight factors,and bundle size.In summary,this thesis addresses the problem of data sparsity,data heterogeneity and extracting item correlations in the bundle recommendation for online supermarkets.A new hybrid rating prediction model using a self-attention mechanism is proposed to fuse item data and calculate item features.The model also utilizes low-rating items to model user preferences and combines user satisfaction coefficients in constructing item graph networks to reveal item relationships.Our approach improves recommendation accuracy and the quality of product combinations by leveraging rating discrepancies.In addition,the proposed bundle recommendation method has high practical value for improving user satisfaction and profit level of online supermarkets.