A TV Product Recommendation System Based On User Behavior

In order to help network operators improve user personalized service recommendation,combining with the user's viewing behavior record provided by a radio and television network operating company,this article constructs the viewing behavior feature vector design from multiple dimensions and implements a linear weighted hybrid model recommendation system.We test and analyze the performance of the implemented system on a real data set.The specific research content is as follows:1.We analysis users' behavior data from multiple angles reveals that both the user's activity and the popularity of TV products have a long-tailed distribution;secondly,the relationship between user activity and the popularity of TV products is analyzed and high-activity users are found.Like on-demand broadcast television products,and low-activity users like to click on the characteristics of popular TV products,and then in-depth analysis of the calculation of user preferences to achieve the key issues of user's item personalized recommendation and the hierarchical relationship between the problems,and we design a recommendation system framework based on user behavior.2.In order to illustrate that different users' viewing behaviors reflect different levels of user's viewing preferences,a simple analysis of the user's viewing behavior is performed.Second,from three different dimensions,the user's behavioral preference is analyzed,and a viewing behavior based on viewing behavior is constructed.The eigenvectors,the eigenvectors based on the on-demand behavior(user_level_directory_feature and user_keyword_features)and the time eigenvectors viewed by the user are constructed.In order to reduce the weight of the popular vocabulary in the user keyword list,the basic idea of TF-IDF is used to construct the users'_key-words feature,to improve the personalized description of the user.3.Due to the phenomenon of so-called "new users" and "new items" in data sets,it is a cold start problem for classical user-based and item-based collaborative filtering algorithms.Classical recommendation algorithms show different performances in different scenarios.In order to improve the accuracy and robustness of algorithms,a hybrid recommendation algorithm model based on weighted fusion is proposed;due to the collaborative filtering approach and similarity based on the user's topic,one of the core steps of the recommendation model is similarity calculation,so different similarity calculation methods need to be analyzed;then,five different recommendation algorithms are designed and implemented;finally,to improve the accuracy of the algorithm recommendation,the five algorithms are combined by linear fusion to form a weighted mixed recommendation model.4.In order to fully evaluate the performance of the system and avoid the over-fitting linearity of the model,the cross-validation method was used to divide the data set first.Secondly,each sub-model was tested from the four evaluation indexes of accuracy,recall,coverage,and popularity.The performance of the user's viewing behavior data set,and the parameters of each sub-model;Finally,based on the linear regression method to give each sub-model a specific weight,and Comparing the performance of the model and sub-model test datasets,it was found that the weighted-mixed model was superior to the sub-recommendation model in accuracy,recall,and coverage.