Research And Application Of Recommendation Model On Highway Freight Details Page

In recent years,the rapid development of the logistics field has penetrated into every aspect of people's lives.In the intelligent life of the internet era,the internet of Things is gradually coming.Therefore,the research on the Internet of Things industry chain is gradually unfolding.Among them,the development of highway logistics is particularly prominent.In 2018,the freight volume of highway logistics accounts for more than 75%of the whole logistics system.With the help of highway freight dispatching platform information,the driver enters the cargo details page after searching information.The recommendation list displayed by the traditional recommendation logic has low accuracy,high decision-making cost of drivers in choosing goods information,long decision-making time and serious empty driving situation of trucks.With the rapid development of artificial intelligence and cloud computing technology in recent years,more and more e-commerce enterprises have benefited greatly from the research of recommendation models in their product display details pages.At present,the road freight dispatching platform has not yet made significant research results in this respect.Therefore,the construction of an efficient and highly stable cargo detail page recommendation model is crucial to the development of the logistics field.Based on the research of traditional recommendation algorithm,this paper divides the whole recommendation model into two parts:the recall model and the scoring model.Among them,the recall model recalls k pieces of cargo information most relevant to the target goods from millions of real-time cargo information,and the scoring model is the probability value that the k pieces of cargo information that is retrieved will be clicked.This paper builds a model based on the real data information of the freight dispatching platform,and the validity and rationality of the real online effect evaluation model.In the recall model,the recall results are obtained by collaborative filtering algorithm and improved collaborative filtering algorithm respectively.The score model obtains the probability of goods being clicked by logistic regression and Xg-boost(Extreme Gradient Boosting).Based on the fitting ability of the off-line evaluation model of Area Under Curve,the authenticity and generalization of the model effect are evaluated on-line based on A/B test.In the recall model stage,the improved collaborative filtering algorithm can get higher AUC value for goods recommendation.In the scoring model stage,data features were screened from 19 to 11 and then to 6 by chi-square test and recursive feature elimination.The AUC value of the improved collaborative filtering algorithm + Xg-boost model is 0.75,which is higher than that of other models.However,Xg-boost's on-line operation efficiency is not high and its interpretability is not strong.The real on-line operation model adopts improved collaborative filtering algorithm + logical regression.The off-line AUC value is 0.65,and the online click-through rate increases by 9%every day.The results of the model achieve the research objectives and bring important economic value to the freight dispatching platform.