Research On Several Issues Of Recommendation Based On Heterogeneous Information

Recommender system is an effective tool to solve the problem of“Information Overload”,with the widely used of it in news,music,social platforms,e-commerce,and other fields to meet the personalized needs of users,personalized recommendations have received more and more attention.As the most widely used personalized recom-mendation algorithm,the traditional collaborative filtering recommendations（CF）only use the user’s historical feedback data to make recommendations,thus having prob-lems such as sparsity and cold-start.Therefore,more and more studies consider mul-tiple types of data to build hybrid recommendations thus improving the accuracy of recommendations,such as social networks,text,images,etc.The data is different in structures and heterogeneous in properties,so it is also generally called heterogeneous information.The heterogeneous information mainly includes structured data（eg,so-cial networks,knowledge graphs,item categories,etc.）and unstructured data（eg,text,images,videos,etc.）,which widely exist in various search engines,social applications,news reading,and e-commerce platforms.They represent the personalized attributes of users and items from multiple dimensions and fields,and deeply mining this hetero-geneous information can help the recommendation to better analyze user preferences and effectively solve the problems such as cold-start and sparsity,which is of great sig-nificance to improve the performance of recommendations.Based on this,we mainly study from the following two aspects:1:Recommendations based on structured data.Structured data generally reflects the complex structural relationship between user-user,item-item,and user-item,such as category information of movies,songs,and products,social networks,knowledge graphs,etc.According to these structural relationships,problems such as cold start can be effectively alleviated.For example,the user’s preference can be estimated through the friend’s preference by using social networks,thus solving the user cold-start prob-lem.However,existing research on the mining of these structured data is still insuffi-cient,therefore more advanced methods are needed to deeply mine the complex rela-tionships from structured data and integrate them into the recommendation framework,thus learning representations of users and items more effectively,and further improving the performance of recommendations.2:Recommendations based on unstructured data.Unstructured data includes in-formation such as text,images,and videos,These data contain rich user preference and item attribute information,which can effectively improve the sparsity problem for rec-ommendations.Existing studies mainly use deep learning methods to learn features of users and items from unstructured data,but these models still needed to be improved and optimized.How to design effective deep learning methods to deeply mine unstruc-tured data,and how to build user and item representations more efficiently according to unstructured data features,are both main issues that need to be studied.Among the above two types of data,social networks（structured data）and review texts（unstructured data）are currently the most widely studied.These two kinds of data are popular in various e-commerce and social platforms and play an important role in alleviating the sparsity and cold-start problems of recommendations.Therefore in the paper,we build recommendation models using the two types of data to improve the performance of the model in the Top-N recommendation and rating prediction tasks,and further solve the problems of cold start and sparsity problems.There are three main challenges in the current research on recommendations based on these two types of data:1.There is a lack of in-depth research on complex relationships in social networks.Most of the current latent factor models（LMF）based on social networks use matrix factorization-based methods such as Collaborative Matrix Factorization(CMF^[1])to integrate user features mined from social networks into the CF recommendation frame-work.Although the learning ability of traditional collaborative filtering methods for latent user features is strengthened through social networks,these methods still can’t fully mine social networks to solve the sparsity and noisy data of social networks,which limits the ability of social networks to improve the performance of recommendations.2.The reviews cannot be fully utilized to alleviate the sparsity problem.Most of the current review-based methods use deep learning methods such as convolutional neural network（CNN）and attention mechanism to learn review features from user re-views and item reviews respectively,then encode user and item latent features using review features directly.However,The way these models construct user and item la-tent features through review texts is too simple,ignoring the interactions between the features of users,items,and reviews.Therefore,the method of constructing user and item features through reviews is not effective,resulting in the model relying too much on the semantic features of reviews.When the number of reviews is sparse,the model can’t further improve the sparsity problem.3.Lack of fine-grained and personalized learning of review features.Most of the deep learning models based on review text learn the review features at three levels:Word-Level,Review-Level,and Document-level.Among these methods,There are still few studies focusing on the Word-Level.In addition,these studies only consider the semantic features of review texts to learn user and item features and ignore the word and review features based on the personalized attributes of users and items.Therefore,the model does not perform fine-grained mining of the review text,which limits the ability of models to predict ratings more accurately using review texts.In response to the above three problems,we first study the recommendation based on social networks to mine social networks more deeply?then,from the perspective of the framework,we study the recommendation based on review texts,aiming to con-struct better user and item feature representations utilizing reviews?finally,from the perspective of mining granularity,we study on how to mine review texts features in a more fine-grained way.The main contributions of this paper are as follows:1.For the first problem,we propose a joint factorization sequential model based on social networks,which is called as Joint Personalized Markov Chains-Based Recom-mendation Model（JPMC）.We first utilize network representation learning（NRL）meth-ods to pre-train the social network,fully mines the user connections in social networks.We then take into account the long-term and short-term preferences of users influenced by social networks and build a joint learning framework to solve the problem of user cold-start in implicit feedback recommendation better and improves the accuracy of the Top-N recommendation.2.For the second question,we propose an interactive recommendation model based on reviews,which is called as Attributed Graph Convolutional Network-Based Recommendation Model（AGCR）.We first use reviews and historical rating scores as the interactive attributes of users and items?We then integrate both reviews and histori-cal rating scores into the representations of users and items based on graph convolutional network（GCN）to learn better user and item features,the model can better construct user and item feature representations by learning a variety of interactions when the number of reviews is sparse,which further improves the sparsity problem and the rating pre-diction accuracy utilizing review texts.3.For the third question,we propose a recommendation model based on personal-ized review features,which is called as User-Specific Satisfaction-Aware Recommenda-tion Model（USR）.We first propose a new concept called personalized user satisfaction,which means different users show different satisfactions using the same word or similar reviews?We then learn review features based on the personalized user satisfaction in both word-level and review-level,which makes the model not only relies on the seman-tic representation of reviews,but also learns personalized review features based on user satisfaction and mines user and item information from reviews in a more fine-grained way,and further improves the accuracy of the rating prediction.