Research And Application Of Recommendation Algorithm Based On User Behavior Sequence

In big data era,information is generated faster and faster,and the amount of data accumulated by all walks of life is also increasing.For instance,in E-commerce scenario such as Taobao and JD,both the number of users and the number of goods are measured in billion.Therefore,it is extremely difficult for an ordinary user to quickly get the goods from the massive commodity pool without any tools.In this case,the emergence of recommendation system has become a necessity.And recommendation system can help users quickly obtain effective information from a large amount of information.After the rise of deep learning,the factorization machine based on deep learning has become a mainstream research direction in the field of click through rate prediction and recommendation system.However,the focus of this kind of model is basically the interaction of features.With the development of the research on recommendation system,more and more researchers start to pay attention to the user behavior sequence and other related information in the sequential recommendation scenario.The user's behavior sequence can be used to represent this user's interest features,and user's interest features has the characteristics of "diversity" and "dynamic change".Therefore,modeling based on the relevant information of user behavior sequence in sequential recommendation scenario can effectively improve the accuracy of recommendation.For example,Alibaba group first proposed Deep Interest Network for sequential recommendation scene and achieved good results in practice in its own e-commerce system.From then on,Alibaba group has successively put forward many excellent sequential recommendation models based on Deep Interest Network,and those models are most modeled based on the user's behavior sequence.Although Alibaba and other groups have made great achievements in the research of sequential recommendation,there are still many problems and difficulties.For example,when the user's behavior sequence is too long,how to quickly and accurately extract the user's interest features is a big difficulty.At the same time,how to fuse the factorization algorithm in the sequential recommendation models based on user behavior sequence is also a direction worthy of research.Based on the above content,this paper improves and optimizes the existing sequential recommendation models,puts forward two excellent sequential recommendation models,and integrates them into the telecommunication packages recommendation system.Based on the theory of sequential recommendation algorithm for user behavior sequence,this paper studied in practice and carried out the following work:(1)Based on Ali's Deep Interest Network,a novel sequential recommendation model named DAMIN is proposed in this paper.Firstly,the calculation method of attention in the user's dynamic interest feature layer is changed,and the reciprocal of the square of the Euclidean distance between the two item vectors is used as its attention weight.Compared with the attention calculation method in Deep Interest Network,this calculation method is more interpretable and achieves good results.Secondly,in order to make full use of the information of the candidate item vector,this paper also weights the candidate item vector,and uses the weighted candidate item vector to sum with the weighted item vector clicked by user.The purpose is to make two similar item vectors more similar and two dissimilar item vectors more dissimilar.Finally,the new weighted vector is used as the input of multi interest feature extraction layer,and capture a variety of user interest features.(2)Similarly,based on the Deep Interest Network,this paper proposes a novel sequential recommendation model named AIN-DIN,whicn integrating factorization algorithm algorithm.Compared with the user's behavior sequence information,the attribute of items belongs to non-sequential feature information.In order to fully mine the relationship between the attribute features of goods,this paper uses the factorization algorithm algorithm to process the attribute features of goods,so as to improve the learning ability of this model.Such as Deep Interest Network,the user's behavior sequence is composed of multiple items browsed or clicked by the user.Therefore,it will be a difficult problem how to fuse the relevant algorithm of factorization algorithm in the sequential recommendation model based on user behavior sequence and explicitly interactive the attribute features of items,and the model names AIN-DIN proposed in this paper can effectively solve this problem.(3)This paper selects the opened and approved datasets of Amazon in recommendation system industry and the telecommunication packages datasets in the real production environment of a telecom company,and carries out a large number of experiments.The results show that the model named DAMIN and model named AIN-DIN proposed in this paper can achieve good performance.Finally,this paper designs and implements a telecommunication packages recommendation system for business analysts based on the actual scenario,and integrates the sequential recommendation models proposed in this paper,which effectively improves the accuracy of packages recommendation.