| In recent years,there has been a surge in the number of information resources on the web,and the question of how to provide users with the most valuable information resources for them has become an urgent one.The recommendation system can filter out the most valuable information for users from the huge amount of data,solving the problem of "information overload" caused by the huge amount of data.The recommendation system must provide the customer with as personalized and accurate a list of recommendations as possible,which must reflect the user’s multifaceted and dynamic preferences over time.The problem of improving the accuracy of recommendation results and capturing user preferences more precisely has become a popular issue in the field of recommendation systems and has led to numerous research results.In response to the above two problems,many mature recommendation algorithms have emerged,and the main problems with recommendation algorithms at this stage can be summarized as the following two problems.Problem one is that each single type of recommendation method has its own advantages and disadvantages,making it a bottleneck in predicting the accuracy of the recommendation results and preventing more accurate recommendations to users.The second issue is that the previous user modelling approach has a singularity problem for modelling user preferences,which cannot be captured more accurately.To this end,this paper first proposes to combine two algorithms,collaborative filtering and sequential recommendation,to form a hybrid recommendation,in order to solve the drawbacks of the existing single method and improve the accuracy of the recommendation results.Secondly,a tensor-based user modelling method is proposed to solve the problem of homogeneity in previous user modelling methods and to improve the accuracy of capturing user preferences.The main work in this paper is as follows.(1)A recommendation model based on a fusion of collaborative filtering and sequential recommendation algorithms is constructed,which integrates the Fastformer model and key-value memory networks to model user rectangles(second-order tensor),enabling the model to capture user features more effectively and to completely characterize user preferences.where the attention mechanism present in the Fastformer model can match weights of different numerical magnitudes to the user’s interaction items and capture information about the items with higher weight in the current sequence,i.e.the main preferences of users in the current historical sequence.The key-value memory network model captures complex and large amounts of sequence information and enables finegrained modelling of each object in the sequence,thereby capturing the fine-grained characteristics of the user.(2)The similarity between the user tensor and the target item vector is calculated by combining the internal and external distances and bias terms that exist between the user tensor and the target item.This calculation method is more flexible than traditional similarity calculation methods and allows the model to achieve greater accuracy in calculating the similarity of the user tensor to the item vector.The input features of this model are divided into four categories: user ID,historical item sequence information clicked by the user,item ID,and bias.The process of modelling the user rectangle is divided into three parts based on the input features: the first part is the modelled user rectangle centroid vector,the second part is the modelled user rectangle offset vector,and the third part is the centroid vector and offset vector together forming the user rectangle.After modelling the user rectangle,the internal and external distances and three bias terms between the user rectangle and the item vector are combined and similarity is calculated with the target item vector,and the user’s preferred items are sorted in descending order according to the calculation results to generate a final recommendation list for the user.The model was experimented on Movie Lens and Ciao DVD datasets with different sparsity showed that the model was able to focus on multiple user preferences and outperformed the baseline method in terms of accuracy of recommendation results,especially in HR and NDCG evaluation metrics by an average of 1.4% and 1.95%respectively over the existing baseline method.In summary,tensor similarity-based recommendation methods involve a variety of key technologies.By analysing the limitations in traditional recommendation methods,this paper constructs a model for using tensor modelling users in a combined algorithm of collaborative filtering and sequential recommendation,The model was also experimentally validated for its excellence in terms of accuracy in fostering recommended outcomes. |
