| Recommendation algorithms play a crucial role in the era of information overload.On the one hand,users can use recommendation algorithms to find things that interest them from the vast amount of information;on the other hand,information providers can use recommendation algorithms to make the produced information gain wide attention.However,the performance of existing recommendation algorithms is often limited by the following problems.(1)Data sparsity problem:The items that users have interacted with usually represent only a small fraction of the total number of items,so most of the elements in the matrix used to record user-item interactions are in a missing state.The data sparsity problem caused by this phenomenon can make it more difficult for the model to model users and items.(2)Inability to adequately learn the characterization problem:The quality of users’reviews varies.Inadequate data cleaning can lead to irrelevant and poorly informed reviews remaining in the dataset;also,complex correlations are often implied between reviews.The representations learned from low-quality,uncorrelated review data cannot accurately represent user preferences and item characteristics.(3)Static learning problem:The topic of all reviews of a user is heterogeneous,i.e.,all reviews of a user usually contain this user’s preferences for different kinds of items,while the topic of all reviews of an item is homogeneous,i.e.,all reviews of an item usually describe the characteristics of the item itself.Such differences in user and item reviews,if users and items modeled by a completely parallel network,would lead to static learning problems.The main research of this thesis to address the above problems is as follows.(1)To address the data sparsity problem,both models proposed in this study consider both rating information and review information.Since users’ text reviews are serialized,semantically rich implicit features that can complement the rating information,recommended by both reviews and ratings can,to a certain extent,alleviate the data sparsity problem caused by using numerical ratings alone.(2)To address the problem of the inability to fully learn representations,this study proposes a local-global perceptual attention model based on the Bert pre-training model,which uses two different levels of attention mechanisms to model users and items:the local attention mechanism can consider the intrinsic connections between reviews;the global attention mechanism can minimize the impact of poor quality reviews on model performance.Experiments show that this model can obtain better performance than models using ratings alone,and also outperforms models using traditional convolutional and recurrent neural networks,obtaining more accurate user and item representations to some extent.(3)To address the static learning problem,the dynamic dual attention model based on the Albert fine-tuning model proposed in this study,based on the local-global perceptual attention model,solves the problem that user features and item features cannot interact with each other due to the use of completely parallel networks in previous models by improving the global attention mechanism,by combining the user’s feature matrix with the item’s feature The attention matrix is computed to dynamically learn the user’s preferences when faced with different items.Experiments show that this model learns more accurate and dynamic representations.Finally,experiments on the Amazon dataset(which contains four subordinate datasets)show that both of the proposed models perform better than the benchmark models(Neural CF,Deep Matrix Factorization,NARRE,DeepCLFM). |
