| With the rapid development of social economy,people’s aesthetic consciousness has also gradually improved.With its profound cultural heritage,rich colors,crystal clear materials,delicate feel and other characteristics,colored glaze is loved by art lovers and young groups,promoting the rapid development of colored glaze products industry.Due to the rapid development of Internet e-commerce,the traditional colored glaze industry has also shifted from offline retail to online sales.In the face of massive data and information on the Internet,how to enable users to quickly and accurately find the content they want is currently the focus of research.The research of recommendation system is to solve this problem.From traditional recommendation algorithm to deep learning model,the research of recommendation system has achieved good results.Compared with the traditional recommendation model,the deep learning model has been applied in various fields due to its stronger data fitting ability and feature mining ability.At present,personalized recommendation algorithms have been used in e-commerce,short videos,social networks,information platforms and many other fields.Through investigation,it is found that most of the recommendation systems on the existing colored glaze products websites are ranked based on the classification of colored glaze products,release time,high score evaluation,etc.In order to enable users to better choose their favorite colored glaze products,improve user shopping experience,enhance user stickiness and loyalty to the platform.Colored glaze products recommendation system based on deep learning is designed and developed.The main work contents are as follows:(1)The recall method based on Graph Embedding is designed,and the user history sequence is obtained by depth-first traversal of the item relationship graph by random walk.Input the generated user history record sequence into the Item2vec model for training to obtain the colored glaze product feature Embedding,and use the inner product distance between the user Embedding and the colored glaze product Embedding to represent the similarity between the user and the colored glaze product.In the process of embedding,the weighted supplementary information is introduced to solve the cold start problem of the system.Finally,the local sensitive hashing method is used to hash map the generated embedding,reducing the number of candidate set similarity calculations.Greatly improve the recall speed and realize rapid recall.And compare the recall algorithm based on Graph Embedding designed in this paper with user-based collaborative filtering recall and item-based collaborative filtering recall on F1 indicators.In each group of experiments,the recall algorithm based on Graph Embedding has better experimental results on F1 indicators.(2)A DIN deep learning ranking model based on attention mechanism is constructed.Using the sequence information of user’s historical behavior,and using the mechanism based on Attention to dynamically build user interest embedding,so that the model can capture user interest.The DIN model is evaluated with the existing deep learning ranking models NeuralCF,Wide&Deep,DeepFM and MLP in AUC,MRR,NDCG@5 and NDCG@10.From the experimental results,it can be seen that DIN model has the best comprehensive performance.(3)Based on the Django framework,a colored glaze product recommendation system is designed and developed.The colored glaze product recommendation system architecture includes :data layer,strategy layer,feedback layer,and application layer,and each layer of the system architecture is designed and implemented.TensorFlow is used to establish model services,and HDFS,Spark,Kafka and other big data storage,transmission and computing frameworks are used to complete feature storage,offline computing and real-time computing.By collecting and processing the user’s historical behavior and real-time features,the offline recommendation and real-time recommendation are completed combined with the recommendation algorithm,and the list of products of interest to users is generated.The system operation and testing can meet the needs of users. |
