| With the development of the Internet and the arrival of the era of mobile Internet,the scale of Internet users continues to expand.More and more people are accustomed to using the Internet to obtain and share information,while the rapid development of the Internet inevitably leads to the rapid growth of user data.The explosive growth of data also makes users suffer from the negative effects of information searching overload.In this context,it is of great significance to help users to filter uninterested contents in the face of massive information,quickly locate services or items that user may interested in,and smoothly obtain effective information,which is conducive to optimizing user experience and improving user's experience.In order to solve this problem,the recommender system emerges at the historic moment.However,the traditional recommender system cannot address the problems of data sparsity caused by the rapid growth of users or items.These problems have a significant impact on the recommendation accuracy,real-time performance and user experience of the recommender system.In recent years,with the development of machine learning and representation learning technology,recommender system has been gradually combined with representation learning.Representation learning converts raw data into a form that can be used by machine learning algorithms.By Combining with representation learning,recommender system not only can use more reasonable representation of data to solve data sparseness problem,but also can learn the topology relation or attribute relation between users and items,so as to enhance the recommendation effect.Therefore,for different recommendation scenarios,fusing different designs of representation learning technologies may bring about a significant improvement in recommendation results.In this paper,we introduce the representation learning technology into two different scenarios of the recommender system,proposes and implements the application usage prediction algorithm SEM based on representation learning and the click-through rate prediction method DSEN based on representation learning.SEM learns the user's usage pattern from the perspective of the user's session,and predicts the sequence of applications that the users are likely to use based on their usage habits.In order to solve the problem of heterogeneity of user sessions,we propose an embedding algorithm for sessions,which embeds the user sessions into the hidden feature space to form a unified feature representation and alleviate the problem of data sparsity.After obtaining an embedded representation of the user's session,we trained a stacked recurrent neural network model to predict the user's application usage sequence.A large number of experimental results based on real data sets show that SEM is superior to the traditional application recommendation methods.DSEN learns the user's potential interest representation from the user's session perspective,and predicts the user's click-through rate for candidate items based on the user's potential interest.We use the session-based embedding algorithm to embed the heterogeneous user session structure into the user session vector of the same dimension.In order to obtain the user's potential interest representation vector,we use encoder and decoder based sequence-to-sequence mapping model to transform the user's session vector sequence into a set of user's potential interest representation sequences.We also designed a dynamic user interest adjustment module,which can adaptively adjust the user interest representation vector according to the different items which are to be predicted.A large number of experimental results based on real data sets show the effectiveness of DSEN in click-through rate prediction. |
PDF Full Text Request