Research And Implementation Of Product Recommendation System Based On Multiple Strategies

The booming of e-commerce provides convenience to users and also brings the problem of information overload.Recommender systems can filter a small number of items of real interest to users and alleviate the information overload problem.In order to balance both effectiveness and performance,the entire recommendation system is generally divided into two phases: recall and ranking.The recall phase uses a simple model to retrieve as many items relevant to the user as possible in linear time to form a recall candidate set,while the sorting phase can use a complex model to accurately predict and sort the recall candidate set in terms of click-through rate to improve the final recommendation effect.In summary,this study constructs a multi-way recall model and a click-through rate prediction model(ACDeep FM)incorporating a self-attentive mechanism to improve the performance and effectiveness of the recommendation service.Finally,a product personalized recommendation system is designed and implemented.This research is practical and important to improve user experience and increase the revenue of merchants and platforms.The research covers the following three main aspects.(1)Multi-way recall model construction.In response to the problem that the traditional recall model has a single recall mode,which leads to a low final recall rate,this paper constructs a multi-way recall model,which uses three different algorithmic models:item-based collaborative filtering recommendation,factor decomposer algorithm and heat-based recommendation to generate a recall candidate set,and then adjusts the weight of each strategy by weighting factors to form a final recall list for use in subsequent models.Experiments show that these algorithmic models designed from different perspectives are fused by weighting to ensure that the recall rate is close to the ideal state.(2)Construction of a product click-through rate prediction model incorporating a self-attention mechanism.Most of the existing click-through rate prediction models ignore the intrinsic correlation between features when modeling the interactions between features to form combined features.In this paper,we propose a product click-through rate prediction model(ACDeep FM)that incorporates a self-attention mechanism.The method first incorporates the self-attention mechanism with adaptive weight assignment to the input features,then splices the output results with those of the compressed interaction network model and deep neural network model in the model,and finally further learns meaningful combined features after subsequent multilayer perceptron layers.Experiments on two public datasets,Ali Tianchi Mobile Recommendation Algorithm Competition dataset and Retailrocket,show that the proposed algorithm improves the AUC by 0.7% and 1.7%,respectively,relative to the very deep factorization machine model,demonstrating the effectiveness of the product click-through rate prediction model incorporating the self-attention mechanism.(3)Implementing a product recommendation system.Combined with the actual user requirements,the multiplex recall model and the product click rate prediction model integrating the self-attention mechanism are applied to the product recommendation system.The system is developed based on Spring Boot and Mybatis framework,mainly including registration and login,user behavior collection and data processing,recommendation results display and other modules.Users can search,browse,collect,add to cart and purchase products in the system.After recording the user's historical behavior,the system uses recommendation models to analyze and process,and finally returns the list of recommended products filtered by the model to the user.