Research On Learning User Shopping Intent And Personalized Recommendation

Recently,with the continuous advancement of computers and information technology,new media such as the Internet and mobile Internet have developed rapidly.Consumers' traditional offline activities,such as shopping,social networking,and reading,are gradually shifting to online.As more and more companies enter the field of e-commerce,the competition in the online retail industry is becoming increasingly fierce.To be able to stand out in an increasingly saturated e-commerce environment,companies try to use a variety of advanced technologies to attract and retain users,such as collaborative filtering,smart recommendations,and machine learning customization.However,the effect of using such a conventional technique is not ideal.Therefore,for academics and practitioners,how to increase online conversion rate is a difficult task.Currently,websites that provide users with recommendations are generally customized on a large scale.Many websites assume that the user's preferences and browsing behavior are static,and then recommend based on the user's history information and preferences.However,the user's intent may change due to the stimuli and information encountered when visiting the website.Conventional recommendations methods(eg,collaborative filtering,content-based data mining)often ignore this dynamic of intent.In addition,conventional recommendation models often regardless of(1)whether the product recommendation satisfies each user's unobserved purchase intent,and(2)whether the recommendation is effective in increasing the purchase conversion rate while reducing the cart abandonment.In response to the above issues,this paper has mainly done the following aspects:First,as foundations of this paper,the related theories and research achievements in user's online behavior and personalized recommendation are reviewed,and the significance of this study is also discussed.Second,user's behavior characteristics are quantified,such as whether the user's current page has accumulated activity,whether it was purchased in the previous session,and so on.The product features are quantified,such as product popularity,whether the product waspreviously purchased.Word-of-mouth stimuli are quantified,such as the number of evaluations,bad reviews,and bad ratings of the product at the corresponding time.The brand's stimuli are quantified,such as the number of brand-views the user has in the session,and the brand's operational enthusiasm.Third,the user's heterogeneity analysis is added to the model,the relationship between the user's demographic variables and their shopping cart selection behavior is also discussed.Fourth,hierarchical Bayesian is used for build observation model,and using hidden Markov chain is used for build the identification model of user real-time intent.The Markov chain Monte Carlo method was used to estimate the posterior distribution of model parameters.It was determined that the user had two states,high and low intent states,and the user shopping cart selection behavior was analyzed under different states.Fifth,in order to verify the validity of the model,the two-state model is compared with the single-state hierarchical Bayesian model.The results show that the user's prediction accuracy under the two-state model is better than the single-state model.