A Study Of Personalized Bundle Recommendation And Its Diversity

Since the birth of the computer Internet,the data that exists on it is changing every second.With the increasing scale of data,how to accurately grasp the preferences of users has become a thorny problem,and the emergence of recommendation systems is a perfect solution to this problem.On this basis,researchers have proposed many effective models and methods.With the popularization and development of the Internet,the "information explosion" has put forward new requirements for recommendation systems.As a special "filtering" tool,it is an urgent problem to solve whether personalized matching can be used to better meet the needs of users and thus improve the effectiveness of recommendations.The most common recommendation method is to use a suitable model or algorithm to recommend the most interesting individual items to users.However,in real applications,users usually interact with multiple items,and service providers want to recommend multiple items as a bundle to gain more revenue.This has given rise to a new recommendation method that recommends multiple items as a whole to the user,a process known as bundle recommendation.Since there is less research on bundle recommendation,there are some important problems that need to be solved urgently:(1)Since the form of bundle recommendation is different from traditional item recommendation,how to add bundles to the recommendation model after characterizing them is the primary problem.(2)Most of the existing bundle combinations are formed by manual participation,and a method is needed to generate bundles automatically and flexibly.(3)The content of bundles is too single,and the generated bundles need to be reasonably optimized.To address these problems,firstly,in the bundle characterization stage,this paper proposes a graph convolutional network-based approach.By constructing a user-itembundle relationship graph,the performance of bundle recommendation is effectively improved by using neighbor information to characterize bundles on the basis of effective characterization of users and items.Secondly,for the generation of bundles,this paper proposes a generation method based on user interest points and a constraint method,which can be flexible to automatically generate bundles and effectively solve the problem of personalized bundle recommendation.Finally,in order to make the content generated by bundles more reasonable,both high matching and diversity,this paper proposes a multi-objective optimization method and a loss optimization algorithm applied to the matching stage.·A GCN-based graph convolutional network characterization method is designed to combine the users who have interacted with the bundle and the items that have affiliation with the bundle into the bundle characterization,which makes the bundle characterization more powerful.·An item relationship graph was constructed based on the user's interaction sequence to generate bundles with a single highest point of interest for the user.In addition,the generated bundles were constrained using a word-frequencyinverse word-frequency approach in natural language processing added to the generation step.·Based on the characteristics of bundles,the generated bundles are reasonably optimized for the characteristics of bundles,and the two optimization objectives of high matching degree and diversity are proposed,and the multiobjective optimization of bundles is carried out to ensure the diversity of bundles.·The positive and negative samples of bundles are added to the loss calculation based on the traditional losses,so that the performance of bundle recommendations can be further improved.