Research On Deep Neural Recommendation Technology Based On Text Information Processing

In the age of big data,recommendation systems play an important role in combating the problem of information overload.Traditional collaborative filtering recommendation algorithms are modeled only based on interactions,which often fail to achieve the best recommendation performance due to the simple and short of data available for model training.Therefore,many works have introduced text as side information into recommendation models to provide additional data input.However,existing text-based recommendation methods often ignore the diversity of user preferences(user preference characteristics will change depending on the item they deal with)or do not make full use of the input text information in the modeling process.To tackle the above problems,with the help of a variety of deep network models,this thesis conducts a research on deep neural recommendation technology based on text information processing,and accomplishes the following research work:(1)To model the diversity of user preferences and item features in recommendation tasks at a finer granularity,this thesis proposes a text-based attentional neural recommendation model,ANAR,which extracts textual feature information from users' and items' reviews through a convolutional neural network and dynamically captures,through a specially designed attention mechanism,the diverse aspect-level attention generated by users in the process of different user-item interactions.Comparative experiments on several publicly available datasets demonstrate that the performance of ANAR's rating prediction recommendation outperforms several state-of-the-art recommendation models,and the designed attention mechanism facilitates ANAR to better model the user-item interaction representation,thus improving the recommendation performance.(2)To make fuller use of the input text information and to better model complex user-item interactions in recommendation systems,this thesis proposes a text-based graph convolutional recommendation model,RAGCN,which models user-item interactions through graph convolutional networks and updates the features of user and item nodes through message passing and message aggregation strategies.The features contained in the items' reviews are extracted via BERT and participate in the whole message passing process as nodes in the graph,which helps RAGCN to better model the feature representation of users and items.Considering that different user nodes and item nodes have different importance to each other,an attention mechanism is designed to regulate the message passing between nodes.Comparative experiments conducted on several publicly available datasets demonstrate that RAGCN has better Top-N recommendation performance compared to several state-of-the-art recommendation models,and prove the effectiveness of the designed attention mechanism.