Research On Sequential Recommendation Method Incorporating Temporal Information

With the rapid development of computer technology and the Internet industry,the information generated in various fields has grown explosively,and information overload has become a major challenge for modern people.A lot of user behavior data is produced as a result of the various ways in which users interact with items.Sequential recommendation can use these behavioral sequences to extract potential user interests and recommend items to them.There are some limitations to traditional sequential recommendation methods.On the one hand,most traditional sequential recommendation methods capture only sequential patterns in user interaction sequences,ignoring the influence of information such as user reviews and time intervals on user interest modeling;on the other hand,some sequential recommendation methods extract users’ multi-grained interests by capturing higher-order associations between sequential items via graph neural networks;however,these methods ignore the fact that multi-grained interests change over time.This thesis proposes a sequential recommendation method incorporating temporal information to address the shortcomings of existing methods and to better utilize the temporal information in item sequences.It uses a variety of deep learning techniques to effectively capture the impact of temporal information on modeling user interest and deeply explore potential user preferences in interaction sequences.This thesis’ s main work is as follows:(1)This thesis proposes a sequential recommendation method incorporating time interval and review text to tackle the problem that some established sequential recommendation methods just use the order of items in the interaction sequence and ignore review information and time features such as time intervals.By incorporating a text feature extraction layer and a temporal dynamic modeling operation into the user preference modeling process,the method improves upon the conventional sequential recommendation method.It is possible to extract user and item features from review documents using the text feature extraction layer,and the temporal dynamic modeling module adds temporal information to the sequential item representation.The above method enables a more accurate and personalized user recommendation system as well as a better representation of users’ short-term preferences.(2)This thesis additionally proposes a sequential recommendation method for user multi-granularity interest drift to address the problem that most sequential recommendation methods cannot capture the user’s multi-grained interest shifts as time passes.The method combines temporal information with graph convolutional networks to construct items in a sequence as a time-aware graph structure.The item relationship graph contains information about the time intervals between nodes,and the item’s embedding in the latent space is learned by the graph convolutional network and attention mechanism in order to capture changes in users’ various fine-grained preferences over time and produce a more accurate representation of the item and recommendation results.To validate the performance of the methods,this thesis conducts comparative experiments on various datasets.The recommendation conclusions are assessed using a variety of evaluation metrics.The experimental results demonstrated that,in comparison to other comparison methods,the two sequential recommendation methods incorporating temporal information have better recommendation effects.