Modeling Long-and Short-Term User Behaviors For Sequential Recommendation

Recommendation system can predict users' potential interest in items,and has been widely used in today's online platform.Sequential recommendation has high practical value in online services(such as e-commerce),so it has attracted more and more researchers' interest.The basic goal of sequential recommendation is to capture the relevance of project transformation.Users' current interests evolve with their historical behavior,which makes it difficult for the platform to make appropriate suggestions.Therefore,it is necessary to describe the evolution pattern of user history and model the sequence pattern of user transactions on the project.Through these user representations,it is easy to recommend appropriate items for each user.Therefore,the main line of research work is to obtain better user representation by using a more powerful sequential model.The user's intention is highly related to the user's interaction with the online platform.Therefore,in the sequential recommendation task,it is necessary to consider the impact of users' short-term intention,users' long-term preference and project characteristics on recommendation.The items are sorted according to the interaction timestamp,and focus on sequential pattern mining to predict the next item that may be interacted.At the same time,it is necessary to explicitly model the interaction timestamp within the framework of time series model,that is,the user interaction sequence should be modeled as a sequence with different time intervals to explore the impact of different time intervals on the prediction of the next project.The main work of this paper is summarized as follows:(1)An attention based deep neural network ADNNet is designed and implemented for sequential recommendation.Considering the long-term and short-term behavior,different modules are modeled to capture the potential interest of users.Convolutional neural network is used to model the short-term behavior sequence of users.The gate recurrent unit module is used to effectively capture users' long-term preferences.In addition,the self-attention module is used to dynamically balance users' long-term interests and short-term needs,and learn the relevant information of users' behavior in the medium,long-term and short-term,rather than a simple combination.(2)We improve the model into a deep neural network TADNNet based on temporal attention mechanism for sequential recommendation.This is a new method to solve the problem of sequential recommendation.The convolution neural network is used to extract the short-term sequence features,and the bidirectional recurrent neural network structure is selected to capture the surrounding event context from two directions,model the overall preference of users,and explicitly model the interactive timestamp using the self-attention mechanism based on time interval to obtain the optimal solution.(3)Through extensive empirical research on six public data sets and baselines related to the latest technology,our model achieves the most advanced performance.The detailed experimental results prove the effectiveness of the model architecture,and finally can successfully capture the long-term and short-term behavior preference patterns in the user sequence,which provides a better solution for the sequential recommendation problem.