Research And Implementation Of Recommendation Algorithm In House Recommendation System

In recent years,the Internet has developed rapidly,and the Internet has greatly enriched people's lives.But the development of the Internet has led to an explosion of information in recent years.With the increase in the amount of information,people now need to spend longer to get the information they want,and it has become increasingly difficult for people to accept and deal with so much cluttered information.And search and recommendation are the main means to solve this information overload.The research focus is mainly on recommendation systems,and the scenario is to allow excess information on the Internet.Solve the problem of excess information on the Internet by using recommended technologies.Provide users with efficient and personalized recommendation services.The recommendation system used here is mainly divided into recall and ranking stages.The recall phase is to find materials that users may find from a large-scale data set.Because the amount of data in this phase is exchanged,we can only use the replaced data and some simple models,and the materials used in the sorting phase are recalled in this phase.What we get is that we use multiple recall channels for robustness and diversity.Then the results of the recall phase are merged and the model is reordered,and these materials are scored through the CTR notice model,and then displayed based on the ordering to the user.Throughout the entire process,the user's historical behavior sequence will provide a lot of information,but we are directly discarding ID-type features when processing features,and using substitution for the information in these sequences has a certain impact on the results we finally pursue.In order to fully benefit the information in the user behavior sequence,certain improvements have been made: 1).First,the embedded vector of the item is obtained in an unsupervised manner based on the user's behavior sequence,and then it is directly converted into the input of the ranking model.2): In order to distinguish the impact of different historical behavior items on the results,we add attention mechanisms to the depth modules of the extensive and in-depth classification models to determine the weight of different items,so that users can be represented by adding sums.Finally,we have done several sets of comparative experiments to prove the effectiveness of these improvements from the accuracy and AUC indicators.