Research On Sequential Recommendation Algorithm Based On Capsule Network

Sequential recommendation generates personalized recommendations for users by learning their static long-term preferences and dynamic short-term preferences.Aiming at the problems in existing sequential recommendations and better generating predictions for users,this paper will start with the following aspects:(1)It is difficult for the existing sequential recommendation methods to generate finer grained short-term preference and complete and detailed long-term preference information for users while considering the sequential pattern.(2)Existing sequential recommendation methods do not consider the features of each item of a sequence that the user pay more attention to,which can determine the interaction of the next item and enhance the representation of sequential patterns.In order to focus on sequential patterns and model finer grained short-term preferences and generate complete user information to express long-term preferences,this paper proposes CCN4 SR.CCN4SR applies horizontal and vertical capsule networks to capture short-term preferences of user,where the horizontal capsule network considers all dimensions of each item in the sequence and constructs it as a capsule to focus on relative sequential location information for each item,vertical capsule network considers a single dimension of each item in the sequence to describe the attribute features of each item to further fine-grained learning user preferences features.For users’ long-term preferences,CCN4 SR applies convolutional neural networks(CNN)and residual connections to learn more abstract and complete features of long-term preferences.In order to construct an embedding matrix that contains more feature information and enhance the representation of sequential patterns by learning the features that users pay attention to in the previous item and interacting with the next item to improve recommendation performance,this paper proposes SPECN.For each item,SPECN applies a self-attention mechanism,using user information as a guide to highlight the features that users pay more attention to in the item,and then concatenates these features behind each item in the original sequence.Secondly,SPECN applies a horizontal capsule network to learn the relative location information and more specific features of each item,these features contain original features and those features that the user pays more attention to of single item or features of adjacent items(features of the previous item that the users pay more attention to and features of the current item)to enhance the sequential patterns features.SPECN uses vertical capsule network to extract the attributes features of each item.In this paper,the relevant experiments are conducted on three real-world datasets:Movielens 1M 、 Gowalla and Amazon CDs,the experimental results illustrate the effectiveness and feasibility of CCN4 SR and SPECN.