Design And Implementation Of Recommendation System Based On User Behavior Seouence And Graph Embedding

With the full popularity of Internet services,recommender systems are becoming an indispensable part of people’s daily life.While enjoying the convenience,people also constantly put forward new demands on the recommendation system.Aiming at the problems of data sparsity,incomplete features,and insufficient modeling of users’ long-term and short-term interests in existing recommendation systems,this thesis conducts research and proposes corresponding algorithms.The main work of this thesis is as follows:(1)Aiming at the ubiquitous problem of sparse target behavior data and the inability of existing graph embedding algorithms to model the hidden high-order semantic information in graphs,a recommendation model based on graph embedding and contrastive learning(GECLR)is proposed.Based on the multi-behavior data between users and products,the model learns node structure information and semantic node information under different views by establishing multiple views,and then introduces a self-supervised comparative learning mechanism to construct positive and negative samples between views,thereby alleviating the problem of sparse supervisory signals.On 3 real data sets,GECLR has improved the NDCG index by up to 24.7%compared with the existing methods.The experimental results show that the GECLR model can model effective high-order semantic relationships.(2)Aiming at the problem that the existing algorithms do not fully consider users’ long-term and short-term interests,a recommendation model based on graph embedding and user dynamic interest(GEDIR)is proposed.The model uses a graph embedding algorithm to establish a multi-behavior view based on all historical interaction behaviors of users,and integrates the node embedding of structural views to model the user’s long-term interest;based on short-term sequence data,Bi-GRU is used to model sequence information.An improved positional encoding is proposed to strengthen the sequential relationship and thus model the user’s shortterm interests.Finally,a context-based weight assignment mechanism is proposed to fuse long-term and short-term interests to obtain the final user embedding representation.On 2 real datasets,GEDIR improves the NDCG index by up to 40.2%compared with the existing methods.The experimental results show that the GEDIR model can model effective user long-term and short-term interest representation(3)Based on the two recommendation algorithms proposed above,this paper designs and implements a personalized product recommendation system based on the Django framework.The recommendation algorithm of the system uses GECLR as the recall model and GEDIR as the sorting model.In the part of system realization,firstly,starting from the perspective of requirements,the motivation for designing the system is demonstrated;based on this,the various requirements of the system and the corresponding outline design are formalized;after that,each module is further designed in detail design,including three modules:interaction,product recommendation,and database.Then select the suitable programming language to implement the system,HTML is used to build the front-end interactive module of the system,the Django framework based on Python is used to build the product recommendation and database modules of the system.Finally,in order to test the effect of the system from the part and the whole,the function test of each module in the system is carried out.Based on this,the overall effect of the system is tested,including reliability,ease of use,and scalability.