Research On Cold Start Recommendation Method Combined With Social Network

The recommendation method refers to the e-commerce platform recommending information and items of interest to users based on their historical behavioral information.It is widely used on a variety of websites in a number of areas including e-commerce,social media and online video.However,for new websites or new users to a websites,their historical interaction information is unknowable,thus creating a data sparsity problem and are greatly affected by the cold start problem,which in turn affects the final recommendation results.Recent research has shown that incorporating social connections for recommendation can solve these problems,which can effectively use the user’s social connections to compensate for the lack of historical interaction information.However,existing social recommendation methods are faced with such challenges as not being easy to extract users’ social relationships from multiple spatial dimensions and ignoring the implicit social relationships between items.Therefore,it is very meaningful to carry out research on new social recommendation algorithms.The main work of this thesis is as follows:For existing social recommendation methods,it is difficult to extract users’ social relations from multiple spatial dimensions and capture the different preferences of users for items.In this thesis,aiming at trusting social relationships,a Social Recommendation Algorithm Combining Multi-headed Attention and Soft Attention Mechanism(SRMSA)is proposed.The algorithm targets trust-based social relationships and enables the model to assign greater weights to task-related elements through an improved multi-headed attention mechanism.The mechanism has both spatial and weighting properties,assigning different weights to different input data and extracts the user’s preferences from multiple spatial dimensions to focus on the most important information in the current recommendation task.Also,in order to reduce the loss of useful information in the score prediction,the feature vector representation of nodes is enhanced by calculating the similarity between user and item features,thus improving the recommendation performance of the model.To address the problem that existing social recommendation methods only consider the social relationships between users and ignore the implicit social relationships existing between items,a Graph Neural Network Recommendation Algorithm Based on Multi-dimensional Social Relationships(MSGNN)is proposed in this thesis.The algorithm introduces multidimensional social relations can effectively alleviate the data sparsity problem by correcting the cosine similarity,mining the similarity between items using users’ ratings of items,and constructing relationship graphs between users and users,users and items,and items and items.By using the modified cosine similarity,it analyzes the score data of different users on items and mines the similarity between items,and using the graph neural network to build the relationship graphs between users and users,users and items,and items and items.Combined with the multi-headed attention mechanism,it captures different interests of users by virtue of multiple feature representations,and reduces the interference of input information by assigning different weights to different feature representations,thus improving the effectiveness of recommendations.The feature similarity is also combined with the neural networks to achieve better score prediction and recommend the goods or services that best meet the user’s current needs,enabling the scalability of social recommendation.In this thesis,we perform experimental validation on two real datasets,Ciao and Epinions,and a cold-start dedicated dataset,Film Trust,the results show that both the Mean Absolute Error(MAE)and Root Mean Square Error(RMSE)have improved compared to the baselines.MAE values of SRMSA and MSGNN in Ciao datasets are 1.67% and 1.46% higher than those of existing benchmark models,respectively,which proves that the proposed method has a good recommendation accuracy.