Research On Heterogeneous Single Class Collaborative Filtering Algorithm

With the development of intelligent recommendation technology,the significance of research on recommendation algorithms for heterogeneous data sources is becoming more and more obvious.Firstly,it is more and more difficult to improve the accuracy of recommendation with purchase data alone.Secondly,the purchase data is generally sparse,and thus requires other data to assist the preference learning task.In this situation,the browse data can be used as a supplement to purchase data,because the former is generally relatively more abundant.For this reason,we focus on studying how to use the purchase data and browse data to establish the recommendation model in order to achieve better performance.There are some semantic uncertainties in browse data.Different browse data are different in expressing the preferences of users.For example,some browse data can express the preferences of users,but some browse data cannot express the preferences of users.It is this semantic uncertainty that makes the recommendation model very difficult to be established.Hence,some studies have focused on how to select browse data with definite semantics in the browse data,but this approach makes the established model more complex,the model training time consumption is also very large.In order to solve this problem,we have proposed two algorithms,i.e.,role-based Bayesian personalized ranking(RBPR)and BPR+.RBPR is modeled from the different roles played by users in the different stages of purchasing goods.By using the data generated by different stages,we train the model separately in order to simulate the role of users in this stage.Besides,we simulate the interaction between roles through the transmission of data.The BPR+ algorithm divides the characteristics of items into buying features and browsing features in the process of matrix decomposition,so that the model can learn more about the user’s browsing preferences and the relationship between purchase and browse.Experiments show that the two aforementioned models are much better than the recommendation model based on the purchase data alone.The model of RBPR is simple,and thus the training speed is faster than BPR+.But the model of BPR+ is more complex than BPR to learn more comprehensive information,so it is usually more accurate.Finally,the complementarity of RBPR and BPR+ algorithms is also studied in an integrated algorithm called EPL.This method makes a unified position representation for items in the recommended list of RBPR and BPR+ algorithms,and then sorts them,and selects Top-K as the final recommendation list.Experiments show that EPL is better than RBPR and BPR+,which indicates that RBPR and BPR+ are complementary to each other.In addition,The study of preference learning algorithm based on feature engineering gives the idea of constructing the features under the recommended scene and how to write the LR algorithm suitable for the recommended scene.Thepreference learning algorithm based on feature engineering can transform the modeling problem of heterogeneous data into feature construction problem,which provides another idea of making recommendation by exploiting heterogeneous data source.