Research And Application Of E-commerce Hybrid Recommendation Algorithm

Recommender algorithms can extract valuable information from massive data and predict users’ future interests.The common recommendation algorithm process can be divided into two stages: recall and sorting.The recall stage is mainly used to quickly select highquality items from a large candidate set of goods,and recommend them to users.The goal is to find as many high-quality items as possible that reflect users’ interests to some extent.The sorting stage focuses on refining the recommendations from the recall stage and providing more detailed sorting and scoring to ultimately recommend items that are more in line with users’ interests and needs.In the sorting stage,typically more complex and sophisticated algorithms are used.The goal is to improve the quality and accuracy of the recommendations,increasing user satisfaction and experience.This thesis constructs a multi-path recall mixed model based on a weighted factor strategy for the recall stage of the hybrid recommendation algorithm in e-commerce.A feature interaction fusion deep neural network model is constructed for the click-through rate prediction model in the sorting stage of the hybrid recommendation algorithm in e-commerce.The mixed recommendation algorithm improves the accuracy and fairness of the recommendation service.The research mainly includes the following two aspects:(1)Construction of a multi-path recall mixed model based on a weighted factor strategy.Since the traditional recall model has a single recall method and low recall rate,and there is a tendency towards the Matthew effect,this thesis constructs a multi-path recall model.The model uses correlation-based collaborative filtering algorithms,item2 vec algorithms,deep walk algorithms,and node2 vec algorithms to obtain recall results,and then uses weighted factors to adjust the weights of each algorithm’s recall results to form the final recall results.Experimental results show that the four different structures of recall models,after fusion with a weighted factor strategy,have significantly improved recall performance.The top 20 recall rate of the multi-path recall mixed model is 0.0748,and the coverage is 0.6073.Compared with a single model,the recall rate has increased by 3.1%,and the coverage has increased by9.9%.(2)Constructing a deep neural network model for feature interaction fusion.To improve the model’s ability to learn higher-order interaction features,this thesis selects the co-action network(CAN),and integrates it with Deep Neural Network(DNN)and Deep Interest Network(DIN).The model explores the deeper relationship between users and goods,simulates the user-item interactions in real-world scenarios,and obtains the click-through rate prediction model in the sorting stage.Experimental results on the public Amazon Review Data Luxury Beauty dataset show that the CAN-DIN algorithm has not significantly increased the parameter count compared to the CAN-DNN algorithm,and the AUC value is higher.Compared with the classic model DNN,the AUC of CAN-DIN has increased by 4.1%,and the accuracy has increased by 5.19%.Compared with the popular model DIN,the AUC of CAN-DIN has increased by 1.7%,and the accuracy has increased by 1.7%,demonstrating the effectiveness of the CAN-DIN algorithm.By finding the optimal hyperparameter combination through experimentation,the final CAN-DIN model’s AUC reaches 0.8699,and accuracy reaches 0.7822.