Research On Sequence Recommendation Based On Feature Extraction Of Gated Graph Neural Network

Information consists of pieces of data,and the sequential relationship between the data can express different information content.The data whose order relationship determines the expression of information is collectively referred to as sequence data.On the Internet,users browse products,click products,compare products,and finally purchase products through the Internet to complete consumption behaviors.These operations of users on products are recorded as interactive data.By analyzing user interaction data,users' preferences can be mined,and user consumption suggestions can be given to users according to their preferences.Interaction data share a common characteristic: sparsity.Data sparsity mainly comes from two aspects.On the one hand,the number of users is small,and the two are in a one-to-many relationship between items;On the other hand,the user's historical interaction records are incomplete,and the reasons at the user level are analyzed: the user does not log in to browse or register a new user,and the retained user behavior information is not enough for preference analysis.The large volume of commodity data,many types,and incomplete user information are the main reasons for the sparsity of interactive data.Dividing interactive data into sequences is necessary.This paper studies the problem of sequence data recommendation.This paper adopts gated graph neural network(Graph Neural Networks with GRU,GGNN)for feature embedding,and studies the sequence recommendation system in three stages:In the first stage,a session recommendation model(SISFGRS)based on gated graph neural network for item structure feature extraction is proposed.Using the short sequence(also called session data)divided by sequence data,a directed graph is generated according to the sequence of clicks,and the feature of the graph structure is captured by GGNN.It is proposed that the session graph generated by the user click reflects the current interest and preference of the session maker,and is extracted as the user feature vector;All conversations constitute a commodity(Item)conversion structure diagram,and feature extraction is performed on the structure diagram to reflect the Item's feature vector.The user feature vector and item feature vector are extracted with the help of gated graph neural network,which solves the problem of anonymity caused by insufficient user information in session recommendation.The second stage proposes the optimization of GGNN sequence recommendation based on knowledge graph structure(KGQRS).The Item feature vector extracted and calculated by GGNN is stored and regarded as the terminal query result(tail entity);The user feature vector trained by the first stage method is used as the index address(intermediate category attribute),and this stage is called the user category;Newly input sequence data is treated as a query vector(head entity).The entire sequence recommendation process can be described as the capture and analysis of the click sequence of the new user,the similarity calculation with the original user category(user feature vector),and the matching query in the storage matrix to find the favorite item corresponding to the similar user as the recommendation result.return value.The process of generating recommendation results and turning them into matching queries solves the cold-start problem of recommendation,and the pre-storage function improves the speed of recommendation calculation.The third stage proposes a distance similarity optimization algorithm(SIMDGRS)based on GGNN feature extraction.The classical similarity calculation cannot satisfy the multi-dimensional similarity of sparse data feature vectors in direction and size,so the spatial distance similarity optimization algorithm is introduced.The feature vector extracted by the gated neural network.The multi-dimensional similarity of distance and direction is synthesized.The above three stages of research on sequence recommendation use GGNN for feature extraction.In this paper,the proposed optimization method is used on real data sets to verify the proposed three researches on sequence recommendation.The experimental results all prove that The proposed method is effective in improving the performance of sequence recommendation.