Research On Sequential Recommendation Algorithms Based On User Preferences

In recent years,with the rapid development of information technology,the variety of information that people can access from the Internet has grown rapidly.Although people can easily access all kinds of information,it has become more and more difficult to search for information that interests them in the huge stream of quantity.Personalized recommendation system can alleviate such problems and help people get the content that meets their preferences from the huge amount of information,and also bring considerable revenue to the enterprise platform.At present,using deep learning to construct recommendation algorithms has become mainstream,among which using deep learning methods to extract user preference information from users' historical behavioral sequence data is the focus of many researches.The existing deep learning-based sequence recommendation algorithms are often not comprehensive enough for user preference extraction,ignoring the problem of the influence of users' own long and short-term preferences.In addition,how to incorporate user rating information and time interval information in user behavior sequences into user preferences is also a key issue.In this paper,we propose a sequential recommendation model based on Transformer encoder and CNN to address the above problems,and incorporate the propagation information and time interval information of popular items on the basis of this model.The research work and the innovation points of this paper mainly include the following aspects.(1)To address the problem that the existing sequential recommendation models ignore the user's own long and short-term preferences,we propose a sequential recommendation model SRec-TC based on Transformer encoder and CNN.After that,the long-term preferences of users are captured by the attention mechanism in Transformer encoder,and then the embedding vector is viewed as an "image" and the short-term preferences of users are extracted by convolutional network,and finally the long-term and short-term interests of users are merged to obtain deeper preference information of users.(2)To address the problems that existing sequential recommendation models ignore the influence of propagation of popular items and the influence of time interval between behaviors,we propose a sequential Recommendation Model SRec-IPTI based on the SRec-TC model,which considers the influence of propagation and time interval of popular items.The model first calculates the propagation power of the items involved in the user's behavioral sequence through the information propagation model,corrects the user's rating information through the propagation power of the items and constructs a user rating matrix after correcting the rating,then constructs a time interval matrix based on the time stamps in the user's behavioral sequence,and then integrates the rating matrix and time interval matrix into the Transformer encoder to obtain a rating and time interval based Transformer encoder.Finally,the encoder and the convolution are used to extract the long-term and short-term interests of the users respectively to obtain more comprehensive preference information of the users.(3)The above models are experimentally analyzed and based on which a video game recommendation system is designed and developed.In this paper,we conduct comparison experiments with common sequential recommendation models on three real and differently sparse open-source datasets.The experimental results illustrate that the proposed SRec-TC model and SRec-IPTI model have good recommendation effects.Finally,the above model is applied to the video game recommendation system to provide recommendation service for the video game information platform.