Research On Sequential Recommendation Method And Benchmark Based On Hyperbolic Space

Recommendation system is widely used in people’s production and life.It has played an important role in information filtering and convenient service in the era of information explosion.Sequential recommendation is an important field of recommendation system,which is widely used in film,e-commerce,short video and other industries.Its main task is to capture the user’s most recent preferences by analyzing the interaction sequence between users and projects,so as to predict the user’s next possible interaction project.The key to the success of recommendation system is the learn representation of user preferences and item characteristics.Many widely used recommendation models are based on Euclidean space,that is,using inner product or Euclidean distance to calculate the similarity between user representation and recommended item representation.With the explosive growth of high-dimensional data in the era of big data,the recommended item information in the recommendation system usually has the characteristic information of hierarchical structure.For example,clothing goods are divided into categories such as tops,which also contain sanitary clothes.Traditional recommendation model based on Euclidean space will produce large distortion when embedding scale-free and hierarchical data representation,which affects the effect of recommendation.In order to reduce distortion,it is necessary to increase the scale of computing space,but with the expansion of space scale,the computing resources required by the recommended algorithm also increase.In view of the distortion of big data recommendation effect exposed in the sequential recommendation method,this thesis uses hyperbolic space to optimize the traditional sequential recommendation algorithm,designs a new sequential recommendation model,and gives a series of benchmark tests.Hyperbolic space is a non Euclidean space with negative curvature constant.It can be regarded as a continuous form of tree,which is more suitable for modeling hierarchy in essence.Aiming at the problem that Euclidean space cannot capture the hierarchical structure of data,this thesis proposes a sequence recommendation algorithm based on hyperbolic space.This method combines factor decomposition machine and depth neural network,and adds translation space to enhance the representation ability while capturing low-order and high-order features;the presentation model is embedded into hyperbolic space to capture the hierarchical information in the data and improve the effect of recommendation.Although hyperbolic space and hyperbolic embedding have received extensive attention in the recommendation system,there is no specific benchmark problem.This thesis gives the performance analysis of recommendation algorithm based on hyperbolic space,and provides the theoretical analysis and empirical results of when to use hyperbolic space and hyperbolic embedding in the recommendation system.In order to test the effect of sequential recommendation model based on hyperbolic space,a large number of experiments are carried out on three real-world public data sets such as Movielens.The results show that this method has better performance than the baselines,and verify the rationality and effectiveness of sequential recommendation based on hyperbolic space.In order to make a comprehensive performance analysis of the recommendation algorithm based on hyperbolic space,this thesis selects five models commonly used in different recommendation fields and carries out experiments on six real-world public data sets such as Epinions.By comparing the performance of each model in Euclidean space and hyperbolic space under different conditions,the final conclusions are as follows:(1)distance model is more suitable for learning hyperbolic embedding than projection model;(2)when the density of data set is small,hyperbolic space is stronger than Euclidean space;(3)when the latent space dimension is small,the performance of the model in hyperbolic space is better than that in Euclidean space.