Research On O2O Hybrid Recommendation Based On User Behavior Preference

In recent years,with the rapid development of mobile Internet technology and the acceleration of people’s pace of life,O2O mode has become an indispensable consumption mode in people’s daily life.Personalized recommendation technology that plays a very important role helps users to filter out information that users love and improves the satisfaction of user experience.Every single recommendation model has different model structures,uses different data,and solves different problems.In addition,it can not effectively mine the user behavior preferences because of the user behavior data is usually sparse.This paper proposes a hybrid recommendation strategy based on user behavior preference,which divides the recommendation process into two stages:matching layer and ranking layer.The matching layer adopts a variety of matching strategies to screen out businesses that users may like from a large amount of business information.These matching strategies include matrix decomposition recommendation algorithm,user based collaborative filtering recommendation algorithm,and recommendation algorithm based on content similarity.The ranking layer adopts the deep learning recommendation model to sort the list of candidate businesses according to users’interests and preferences to obtain the top-N recommendation list.Through the experimental analysis on the user rating behavior data of a review website,the hybrid recommendation algorithm based on user behavior preference is similar to the average recommendation effect of multiple single model recommendation algorithms in these indicators including accuracy,recall,FI and popularity,and the coverage index is significantly higher than the average recommendation effect.This paper shows that the hybrid recommendation model based on user behavior preference combines all the advantages of multiple single recommendation models.On the one hand,it can not only provide better comprehensive recommendation effect combined with group intelligence and personalized preference,but also avoid the problem of poor recommendation accuracy for users with sparse behavior data.On the other hand,it can also mine some long tail businesses,which improves the novelty of the recommendation.