A Research Of Cross-domain Recommendation Algorithm Based On Transfer Learning

With the third wave of information technology,the data overload is becoming more and more serious.Nowadays,the internet data is growing explosively,and online users can’t describe their requirements accurately.However,Recommendation systems can provide personalized recommendation services for each user by analyzing the interaction behaviors and context information between users and items,which can help users process information more and more effectively.Recommendation algorithm is the core technology of recommendation system,the main task of the algorithm is to predict the user’s rating or click rate.The existing recommendation algorithm mainly study the user’s rating or click rate of item in a single domain,modeling is done using only information such as characteristic interactions,user behaviors,and context within a single domain.However,Applications often need to deal with complex data from multiple modules in different domains.By analyzing users’ behavior logs,the recommendation system can obtain users’ interaction data from multiple domains,which can comprehensively reflect the multi-dimensional interests of users and also provide data support for the implementation of cross-domain recommendation(CDR)algorithms.This paper combined the ideas of transfer learning and deep learning,and then found a better solution to the cold start problem and data sparsity problem.Based on existing research of single-domain recommendation algorithms,this paper mainly studied the cross-domain CTR prediction problem.The key of CDR recommendation algorithm is the knowledge transfer between multiple domains.This paper designed and implemented a cross-domain prediction model named MANet,which can better retrieve the knowledge from source domain data.The primary job in this paper are as follows:1.This paper comprehensively combed and summarized the evolution process of traditional recommendation algorithms and deep learning recommendation algorithms,and introduced the key technologies and main ideas of cross-domain recommendation algorithms.2.Based on the idea of deep learning and transfer learning,a cross-domain prediction model is proposed.This model comprehensively considers three kinds of user interests from different domains,cross-domain long-term interest,source domain short-term interest and target domain short-term interest.The high-dimensional discrete feature data such as user characteristics,source domain historical behavior and target domain historical behavior are densified by feature embedding technology.The pre-training model can accelerate the convergence speed of the master model and provide support for subsequent processing of multi-dimensional heterogeneous data.3.The attention mechanism is optimized in the process of cross-domain knowledge transfer.Thus,the function of dynamically extracting effective information from the nearest marked items in the source and target domain and adaptively fusing different interest representations is realized.Also,the algorithm realized the function of adaptively adjusting the interests weight of different users,so as to skillfully complete the migration and fusion of users ’ long-term interest and short-term interest.4.The attention network and the fully connected network are constructed,and the interest vector after migration and fusion is used as the original input of the neural network.The fully connected layer is used to complete the automatic crossover task of all features.The comprehensive use of multi-domain data not only improves the prediction performance of the target domain and the utilization of network resources,but also makes the recommendation results richer and novel,which is helpful to break the information cocoon.5.Multiple sets of offline experiments are designed to compare the MANet model proposed in this paper with lots of classical single-domain and CDR algorithms,and the effectiveness of the three interests fusion modeling and hierarchical attention mechanism of the MANet algorithm is verified on the data set containing a large number of samples.