Research On Recommendation Algorithms By Incorporating Multi-criteria Ratings And Multi-source Texts

With the rapid development of Internet,people encounter the problem of information overload despite enjoying the convenience.Recommendation technology is an effective method to solve the problem of information overload,which has become an indispensable function of all kinds of network platforms.For example,shopping websites recommend potentially interesting itemsto users;movie websites help users find favorite movies more quickly and efficiently;and tourism websites push the appropriate hotels according to users'traveling needs.Therefore,personalized recommendation has become one of the most important techniques of web platform intelligence.Collaborative Filtering is a recommendation technique which has been studied in depth and widely used.Most existing CF-based methods only rely on user's historical behavior data of a single perspective,that is,an overall rating,to make recommendations.However,it is hard to depict users' detailed preferences by simply making use of user's overall rating,and these methods are not good at solving the problems of data sparsity,extensibility,interpretability,accuracy and cold-start,etc.With the increase of the types and the sources of information,more and more user behavior data are available for the recommender system,such as multi-criteria ratings,review text of users and descriptive information of items.Such behavior data bring a new opportunity to solve the above-mentioned problems existing in the recommender system.In the meantime,how to integrate such multi-source heterogeneous information in a unified and extensible framework in order to improve the recommendation accuracy and efficiency is also an important problem to be solved.In this paper,we focus on the problem of data sparsity,extensibility,preference dynamics,cold-start and data heterogeneity,and study the recommendation algorithm that integrates multi-source information based on existing work in the field of multi-criteria recommendation,sentiment analysis recommendation and content-based recommendation,etc.Our main contributions are as follows:1.We propose a multi-criteria recommendation approach based on users' rating patterns and items' rated patternsIn the case of multi-criteria ratings,the problem of data sparsity becomes more serious than in the case of single overall ratings,which results in the fact that it is more difficult to find the users who have similar ratings with the target user in terms of multiple criteria.In order to solve this problem,users' rating patterns and items' rated patterns on multiple criteria are first extracted from user-item-criteria matrix based on codebook algorithm.Then,users with similar rating patterns and items with similar rated patterns are clustered into the same cluster.Finally,we extend the vector of user-item-criteria with users' rating patterns and items' rated patterns and integrate them into a unified factorization machines(FMs)framework to make recommendations.Due to the fact that FMs use interaction parameters instead of independent parameters to model feature interactions,the proposed algorithm can obtain high accuracy even in the case of data sparsity.Meanwhile,because of its linear computational complexity,the efficiency of our algorithm is relatively high.Experimental results show that our proposed approach can significantly improve the prediction accuracy of multi-criteria recommendation and effectively cope with the data sparsity problem and extensibility problem.2.We propose a dynamic recommendation model based on aspect-level sentiment analysisUser review data implicit the different aspects of items and users' sentiments towards them.Also,users' aspect sentiments vary with time.Most existing methods based on aspect-level sentiment evolution analysis rely on the whole review data,which is only able to find the emotion variation of a group of users instead of that of an individual user.Moreover,these approaches are usually based on review data from the same category,ignoring that items from different categories have different aspects.This makes it impossible to capture the category-aware aspect preferences of users when corpus contains reviews of items from multiple categories.On the other hand,it is difficult for time-aware dynamic modeling approaches to model users' dynamic preferences directly in item item space due to the fact that the number of commodities purchased by a single user is relatively small.In this work,we jointly model the dependency relationship among category,aspect,aspect-sentiment and time in order to find how users' aspect prefcrences vary with time on different category items.Experiment results on two real-world data sets show that our proposed model has better performance than the recommendation methods based on static aspect-level sentiment analysis and time-aware based recommendation algorithms.3.We propose a unified neural network recommendation model by incorporating multi?criteria ratings and user reviewsMost existing methods incorporating ratings and reviews usually only consider an overall ratings and are based on the topic model.These methods mainly have two defects:(1)topic model-based method ignores the word order in reviews.However,word order is key to modeling semantics accurately;(2)they suppose that topic factors are directly equal to the latent factors of users and items,which is not reasonable due to the fact that the topic factors are limited to review text and not every word in the review is related to rating.In this paper,we propose a deep learning recommendation model in which a neural factorization machines-based regression model is used to project the feature interactions between user,item and criteria into ratings,and a gated recurrent neural networks are employed to "translate" the feature representation of user,item and criteria into a review.Extensive experiments on three real-world datasets demonstrate that the proposed model achieves significant improvement over the models incorporating single ratings and user reviews,as well as other related models.4.We propose a local topicalization pairwise convolutional neural network recommendation model by incorporating multi-source textExisting association rules-based recommendation methods have cold-start problem and content-based recommendation methods are usually based on the similarity of attributes in item titles and item descriptions,ignoring the attributes and semantics in unstructured user reviews.This makes it difficult to excavate deep attributes and semantic associations between commodities.To overcome the defects of the methods above,we first extract the topics from user reviews,and then design a pairwise convolutional neural network to model local attributes and semantic associations by incorporating item titles,item descriptions and user reviews.Finally,we make recommendations based on the deep attributes and semantic associations of item.Experimental results on two real-world dataset show that our proposed method outperforms the traditional association rules-based methods and content-based methods.