Research And Implementation Of Product Recommendation Algorithm Based On Graph Embedding And Deep Learning

One of the current research hotspots of recommendation systems is to mine the implicit user interests based on user interaction sequences.While achieving significant results,there are also data sparseness and cold start problems in personalized recommendation scenarios.This subject has conducted in-depth research based on the above issues,as follows:This subject uses graph embedding technology to solve the problem of data sparsity and user goods cold start.Based on the conventional graph embedding algorithm,this thesis introduces the side information of goods.It uses random walk algorithm is used to model all kinds of attributes of goods and get the matching embedding vector.After that,we can further integrate them.It transforms high-dimensional sparse feature vectors into low dimensional dense feature vectors to solve problem of data sparsity.Meanwhile,the side information can effectively solve the problem of cold start.Graph embedding technology can not only solve the problem of sparsity and cold start,but also inputs the pretrained embedding vector into the upper deep neural network.It accelerates the convergence speed of the network and further improve model performance.Based on the Embedding vector earned by pretraining,the Multiheaded self-attention proposed in the Transformer model is introduced into the product recommendation scene,combined with position encoding to model the dependency relationship between users' historical purchases.Afterwards,the AUGRU(GRU with Attentional Update gate)structure is used to filter the user interest evolution path.And a more accurate user interest vector representation is obtained.In this subject,the above technologies are integrated.Besides the CTR prediction model GEARec is proposed for personalized recommendation of products.The GEARec model is constructed by a pretrained Embedding layer,a Multiheaded selfattention network layer,AUGRU and MLP.After the input user behavior sequence Embedding vector is calculated by the Multiheaded self-attention network layer and AUGRU.Then result is spliced with other Embedding,and then input into the MLP to automatically learn the nonlinear relationship between the features.Finally,it outputs the probability of user buying candidate product.The GEARec model constructed in this subject is compared with the mainstream recommendation model in recent years.The experimental shows the GEARec recommendation model on the Books and Electronics data sets.Compared with the current more advanced DIEN model,the RelaImpr index has increased by 2.27% and 2.42%respectively.At the same time,the author makes a more detailed comparative experiment on each module of the GEARec model,so as to explore the core factors that affect the effect of model recommendation,further optimize the model structure and improve the recommendation effect.Finally,this thesis designed and implemented a product recommendation system based on the Spring Boot framework.It integrates the GEARec model with the recommendation system.It also implements RPC interfaces to provide product CTR estimation and Top-N recommendation services.In addition,the system counts and displays the experimental data used by the training model in the form of tables and charts.It provides functions such as data upload and download,model training,and model weight file download.