Rating Prediction Task Study Based On Reviews

We mainly research the review-aware rating prediction of recommender systems.So far the best method of review modeling in rating prediction is Convolutional Neural Networks(CNNs)or its variants.Although CNNs performs excellent in text modeling,it has some drawbacks for review modeling.On the one hand,CNNs cannot capture enough valid-phase features for short reviews,due to its limited information.On the other hand,max pooling operation CNNs fails to capture the frequency information of valid features,this leads to incomplete review modeling especially for long reviews.Hence,our innovation points as follows:◆ We suggest to utilize the word frequency information of reviews,and further propose a model called Deep Latent Dirichlet Allocation(DLDA)based on word frequency information for rating prediction task.Experiments on Amazon datasets illustrate that our DLDA model outperforms the state-of-the-art models based on word frequency information.◆ For the problem of short reviews,we propose a network called Word-Driven and Context-Aware Matrix Factorization(WCMF-PSC).It merges the outputs of DLDA and CNNs by our proposed PSC method.Experiments on Amazon datasets show that WCMF-PSC outperforms the state-of-the-art models for users and items with short reviews on rating prediction task.◆ For the problem of long reviews,we propose Word-Driven and Context-Aware Matrix Factorization based on Netural Networks(WCMF-NN)and Word-Driven and Context-Aware Matrix Factorization based on Control Loss(WCMF-CL)networks.WCMF-NN utilizes neural networks to merge DLDA and CNNs,while WCMF-CL merges the outputs of DLDA and CNNs through adding control loss function.Experiments on Amazon datasets demonstrate that both WCMF-NN and WCMF-CL outperform the state-of-the-art models for users and items with long reviews on rating prediction task.