| In the era of traditional media,the source of information production is short,and the quality and dissemination of information depend on the subjective intention of the information producer,which limits the choice space of the audience.With the popularization of the Internet and the development of information technology,information overload has become another problem for information receivers.As the times have progressed,recommendation algorithms have become the antidote to information overload,and the personalized recommendation of graphic information gradually walks into ours daily life.A good recommendation algorithm can not only improve the transmission efficiency of information and enhance the reading experience of users,but also expand the commercial value of information platform and change the industrial pattern.Up to now,there are still many problems to be solved in the field of information recommendation,such as the conflict between algorithm performance and computer computing power,the quality of recommended information is uneven,the algorithm model is diverse and complex,each algorithm has its own advantages and disadvantages,many algorithms still have a certain distance from practical application and so on.Aiming at the problem of various algorithm models in the field of recommendation,this paper designs and implements a recommendation system based on multiple recall.The core of this system is to design corresponding recommendation algorithms according to the popular recommendation in the industry,which could control the overall recommendation quality by the means of multiple recall.This paper is oriented to the graphic information recommendation system,and mainly realizes and integrates three recommendation algorithms,include ALS,item-based and GBER,which cover matrix decomposition recommendation algorithm,item-based collaborative filtering recommendation algorithm and graph-based neural network related recommendation algorithm.Aiming at the problem of conflict between algorithm performance and computation force,this paper solves the problem of excessive vector dimension in item-based algorithm effectively by improving the computation method of vector similarity.By constructing a new adjacency list storage method,the random walk algorithm in GBER algorithm performs better.In view of the longterm deviation of users' interests,this paper divides the information sequence into buckets in time,which solves the problem of information having timeliness and further improves the performance of the recommendation algorithm.To solve the problem of insufficient utilization of information feature information in recommendation system,Word2 vec algorithm was innovatively combined with decision tree,and feature importance was taken as an important parameter in neural network training to participate into training,which effectively alleviated the calculation pressure and improved the recommendation accuracy of the algorithm.The system has been deployed and is running well online.The data set is taken from the user's interaction behavior with information and the characteristic information of information,which belongs to the APP of Beijing Weli Technology Co.,Ltd.Weli Kan Kan.The algorithms compared in the experiment include ALS,item-based,GBER,the twin tower model and graph autoencoder,and the key evaluation indexes of the experiment include click-through rate which said as CTR,running time,and the diversity of recalled items.Based on the evaluation index results,the following conclusions are drawn in this paper:compared with the previous improvement,the CTR of the GBER algorithm is improved by17.4%,and the CTR is the highest among the online algorithms,effectively improving the accuracy of the recommendation.The diversity of recall information in ALS algorithm is 7.1,which is rich in information types and high in recommended information quality.The Itembased algorithm only takes 8 minutes to run and has a fastest computation speed,but poor algorithm performance. |
PDF Full Text Request