Research On Recommendation Model Combined With Social Information

The traditional recommendation model mainly recommends items that users may need based on user preferences and other information,and plays an important role in alleviating the problem of "information overload".It is widely used in various fields and has attracted the attention of researchers.With the rapid development of social networks and the in-depth integration of social and various applications,a recommendation model that combines social information is one of the more active research directions in the current recommendation research field.A wealth of social information is obtained through social networks,and social information is integrated into it.In the traditional recommendation model,the unknown item score is predicted to generate recommendation results,which greatly alleviates the data sparseness and cold start problems that exist in the traditional recommendation model.However,most existing recommendation models that combine social information have the following problems:1.Only consider a single type of social relationship in social networks,and fail to consider other different types of objects in social networks and the relationship between objects,such as project nodes.,The relationship between the item and its attributes,etc.;2.Usually each user is projected to a point in the space,which is not enough to correctly model the strength of the user-item relationship in the implicit feedback and the heterogeneity of the relationship between different node types So that the two cannot be effectively combined to achieve a two-way interactive effect.In addition,since most social networks are sparse,although social information is a supplement to scoring data,most of them are sparse.It is necessary to consider the double sparsity of social information data and scoring data at the same time.This paper conducts research on the above issues,fully excavates and effectively utilizes the rich social information in social networks,and then improves recommendation performance.The main tasks are as follows:(1)Aiming at the problem of double sparsity of scoring data and social information data.A recommendation model based on metric learning and network representation learning(A Social Recommendation Based on Metric Learning and Network Embedding,SRMN)is proposed.Introduce metric learning to replace the inner product in matrix factorization,use the combination of matrix factorization and metric learning to capture user behavior habits,use distance to reflect user preferences,and add constraints to ensure that users are closer to favorite items and make recommendation results more reliable And it is more explanatory;learn to obtain user social network representation from social network through network representation learning,effectively mining and use the hidden information in the user's social network;use user social network embedding representation and project potential feature vector to interact,use interaction Reflect user preferences,combine distances to reflect user preferences and interactions to reflect user preferences,to achieve the goal of maximizing user preferences,so that scoring data and social information data are more fully combined,and the double sparsity of scoring data and social information data is alleviated,and recommendations are improved performance.Comparative experiments on three real data sets show that the SRMN model has better recommendation performance on double sparse data sets.(2)There are not only single user social relationship types in social networks,but also other different types of objects and relationships between objects.Considering multiple types of relationships can more accurately determine user preferences.In this regard,a recommendation model based on heterogeneous information network(A Social Recommendation Based on Heterogeneous Information Network,SRHM)is proposed.Using heterogeneous information network extraction means learning to extract and use effective information in heterogeneous information networks,such as multiple types of nodes and relationships between nodes,to obtain user and item representations based on heterogeneous information networks;through user-item relationship construction User-item-specific relationship vector,each user is translated into multiple points,combined with user representation based on heterogeneous information network and user-item relationship vector,so that user-to-many mapping can be regarded as one-to-one Mapping,introducing metric learning to solve the limitations of inner product as a scoring function,using Euclidean distance to calculate the distance between users and items,Euclidean distance to measure preferences,effectively modeling the strength and heterogeneity of the user-item relationship in implicit feedback To achieve the purpose of fully mining the heterogeneous information of users and items in social networks,and to improve recommendation performance.Comparative experiments on three real data sets show that the SRHM model has better recommendation performance.