Research On The Recommendation Methods With Representation Learning Based On Heterogeneous Information Network

With the continuous development of mobile Internet technology,the network structure and its topology become increasingly Complicated,and the data scale explosively expands.The tasks how to quickly and accurately collect and obtain more valuable data and its characteristics from the mass information have become the current focus of research.Personalized recommendation algorithms are effective to dig out potential information of user preference by user data on historical activity.And collaborative filtering recommendation algorithm,as a classic personalized recommendation algorithm,can accurately find users preferences and tendencies,efficiently filtrate,predict and recommend the products for customers.However,with the rapid growth of data scale and increasingly complex recommendation problems,collaborative filtering algorithms still cannot avoid such problems as below:the methods only use the matrix interaction between users and items,and it also lacks effective solutions to the sparse user data and cold start problems.In recent years,Heterogeneous Information Network(HIN),composed of multiple attribute nodes or edge connections,can integrate complex multi-source Heterogeneous Information and has strong data modeling and analysis capabilities.It has been widely studied and applied in many complex tasks of big data analysis and mining.Because heterogeneous information network Methods have high application flexibility in solving many complex problems on analyzing and processing the heterogeneity of modeling representational data.combined with a variety of rich auxiliary information,it has also been used by domestic and foreign scholars in the analysis processing and characterization modeling process of automatic recommender system.Most of these heterogeneous recommendation algorithms acquire relevant information of users based on meta-path,and also use network representation learning to carry out representation fusion.With the improvement of recommendation performance,the following problems still exist:traditional collaborative filtering recommendation methods only consider the historical interaction behavior of users and commodities,and cannot take advantage of the complex relationship hidden in the heterogeneous information network,so it has scalability problem.In view of sparse user history information,local inference with heterogeneous auxiliary information may lead to reasoning conflict and sparse inconsistency.Secondly,the recommendation methods based on heterogeneous information network usually only consider the low order interaction in heterogeneous information network,and ignore some of the high order relationships.It is easy to bring the problem of missing information.Finally,users' interests and preferences may change with the passage of time,and interest drift may affect the recommendation effect.In view of these problems,this thesis proposes a series of research methods on Heterogeneous Information Network Recommendation,and makes use of the rich structure and semantic information of Heterogeneous Information Network to make recommendations.The main work contents are as follows:(1)In terms of the scalability problem of traditional collaborative filtering recommendation methods,a heterogeneous information fusion network embedding attention preference recommendation method(MFFHINE)was proposed,which could effectively utilize attribute information through the joint optimization of matrix decomposition model and fusion function.Then,a fusion strategy based on attention mechanism is adopted to organically fuse the preference features generated by different weight element paths to solve the local sparse inconsistency problem in the inference process of heterogeneous information networks.(2)As for the problem of information missing based on heterogeneous information network recommendation method,a high-order recommendation method(RLTE)based on triangular embedding in heterogeneous information network is proposed,which is guided by multi-level embedding.The skip-gram model and the local triangulation structure are used to mine the hidden low-order and high-order relationships in the heterogeneous information network respectively,and the problem of information missing is alleviated by the feature fusion of these potential information.(3)To solve the problem of user interest drift in the process of dynamic recommendation,a strategy method of tracking user interest preference with time sequence factor(DRLTE)is introduced,which combined with triangular high-order and multi-level information fusion to make dynamic recommendation to solve the uncertainty problem of interest drift.Finally,the whole work is implemented in Python program,compared with the relevant benchmark and the state-of-the-art(SOTA)recommendation algorithms and the parameters analysis.Experiments on real large-scale data sets also show that the proposed Methods are effective and feasible.