Integrating Aspect Information Of Review Text Towards Deep Learning Recommendation

With the rapid development of Information technology,the Internet has been popularized in entertainment,consumption,work,learning,and many other fields,affecting various aspects of social life and development.Driven by diversified Internet applications,almost unlimited connectivity,and the rapid development of information sharing mechanism,the scale of Internet users is expanding rapidly.The communication channels of network information are increasing sharply,which may be further expanded shortly.Although the influx of Internet information is considered an essential factor in promoting the rapid development of many application fields,it also increases the cost of information acquisition,which leads to the problem of information overload.The recommendation algorithm models the user interest based on the user’s historical behavior.It takes the user interest as the core to filter the massive online information to help users screen the appropriate content or goods and optimize the user experience.It provides a feasible solution to alleviate this contradiction.The traditional recommendation algorithm is mainly based on the attribute characteristics and scoring of users or items,but the attributes of users or items are mostly static basic characteristics,which can not reflect the dynamic preferences of users in different scenarios.In recent years,the rapid development of online platforms and users’active sharing of various product experiences have promoted the continuous growth of user reviews.More and more research began to pay attention to the review text containing a wealth of information about user-item interaction.However,most review-based recommendation algorithms focus on the document-level or review-level,ignoring the rich and fine-grained aspect-level information in the reviews.In fact,as the description and evaluation of item attributes are involved in user reviews,such as prices and user evaluation descriptions of prices,aspect-level information helps us reveal the more potential detailed user feelings and more profound user experience behind the overall score.It has irreplaceable importance for the recommendation algorithm.Although some work has tried to integrate aspect-level information into the modeling process of recommendation algorithms,most methods only model recommendation algorithms based on aspect-based information.There are few reports on joint modeling multi-source information of aspect features,review features,and ID features.The structure and performance of the model need to be further optimized and improved.In addition,most aspect-based recommendation algorithms need domain-specific aspect-level information as a priori knowledge,limiting the model’s universality to a certain extent.Given the above problems,this thesis mainly carries out the following four aspects:（1）The recommendation performance of the aspect-based sentiment analysis method in the actual task scenario is compared.Based on two mainstream aspect-based sentiment analysis methods,such as the method based on the sentiment dictionary and position-aware tagging scheme,this thesis constructs two basic aspect-based recommendation models Neural CF_AspectSenand Neural CF_AspectJET,verifies the effectiveness of aspect-level information for the recommendation algorithm,and compares the application effect and task adaptability of the two models.Then the aspect-based sentiment analysis method based on the sentiment dictionary is selected as the final aspect-level information extraction model.（2）A deep learning recommendation model ABRLR based on multi-dimensional feature fusion is proposed.This thesis constructs the deep learning recommendation model ABRLR based on multi-dimensional feature fusion by fusing information such as review,aspect,and ID features.By comparing the application effects of two aspect-based sentiment analysis methods based on the sentiment dictionary and position-aware tagging scheme in the framework of the ABRLR,we further verify the effectiveness of the method based on the sentiment dictionary.Meanwhile,this thesis compares the modeling performance of the ABRLR model and four mainstream models through Amazon public review dataset.The experimental results show that compared with the optimal benchmark scheme,the ABRLR model integrating multi-source information can significantly improve the prediction performance of the recommendation model,and the maximum decrease of the MSE index in the three datasets reached 4.64%.（3）The framework of ABRLR model based on multi-dimensional feature fusion is optimized.In the modeling process of the aspect-based recommendation model,this thesis compares the influence of the different fusion modes of various information sources such as review,aspect and ID features,user-item interaction modes,and processing modes of the user-item interaction feature on the effect of the recommendation model.Based on the performance of various schemes,the various information source fusion mode and processing mode of the user-item interaction feature in the ABRLR is adjusted to further optimize the effect of the model.（4）An Deep Learing recommendation optimization model ABNAR based on attention mechanism is proposed.This thesis constructs the Review/Aspect Double-layer Attention Network by integrating review and aspect features.Meanwhile,it proposes the Joint Learning Layer of User and Item for jointly modeling user or item vectors to further model the informativeness of review text and the interrelationship between the user and item vectors.Further,based on the optimized ABRLR model framework,the ABNAR model is proposed by introducing the Review/Aspect Double-layer Attention Network and the Joint Learning Layer of User and Item,which learns the user or item vector more fully and achieves better prediction performance than the ABRLR model.Compared with the ABRLR model,the maximum decrease of the MSE index in the three datasets reached2.30%.