Research On Recommendation Algorithm Based On User Short-term Interest And Multi-objective Optimization

In the era of information explosion,recommendation system can solve the matching problem between users and information flow.On the one hand,it can solve the problem of information overload faced by users;On the other hand,it also creates profit space for enterprises.Therefore,a number of recommendation algorithm models based on big data technology have emerged in industry and academia,which improve and optimize the recommendation system from different angles,including two research branches: user behavior sequence modeling and multitasking learning.The recommendation system can be divided into two stages: data processing and model construction,in which the model construction stage can be divided into modeling for recall layer,sorting layer and rearrangement layer.In the data processing stage,the user behavior sequence is often modeled,that is,the user’s click,purchase,collection and other behavior history data are mined,but the role of the user’s long-term and short-term interest is often ignored,and the long-term and short-term interest plays an important role in depicting the user’s portrait and improving the model prediction ability.In the model construction stage,the traditional recall layer model is often simple in structure and can only model a single task(such as click through rate prediction task),while good recommendation results need to model multiple tasks(such as click through rate,purchase rate,collection rate and other prediction tasks)at the same time.In view of this,this paper makes innovations in data processing and recall layer model construction.In the data processing stage,this paper tests based on wechat video data.Firstly,the user’s short-term interest vector is extracted by the short-term interest attenuation formula,and the influence of different attenuation rates on the short-term interest vector is explored.Then,this paper uses PCA principal component analysis and kmeans clustering to extract the user’s long-term interest vector,and explains the long-term interest vector through statistical analysis.Finally,AUC is used as an index to verify the effectiveness of long-term and short-term interest vector features.In the model construction stage,this paper tests the real-time data stream recommended by Alibaba click and transformation prediction data set and Tencent information stream.In the recall layer,this paper proposes a personalized multi task multi tower model with differentiation for large-scale recommendation(PMD).The model uses a sub model of knowledge distillation and user preference based learning.In the recall layer,the traditional single task model is improved into a multi task model to further cooperate with the multi task learning model in the subsequent sorting layer,Improve the comprehensive prediction effect of the model.In theory,this paper puts forward new solutions to the characteristics of users’ long-term and short-term interest vectors and the construction of multi task recall model,which enriches the methodology of recommendation system.In practice,the effectiveness of the model in the recommendation system is verified by two public data sets and an online data stream,which has important theoretical and practical significance.