Research On Short Video Real-time Recommendation Method Based On Knowledge Graph

In today's mobile Internet era,the multimedia audio-visual entertainment industry has developed rapidly,and the entertainment information shows an explosive growth trend.The problems of information overload and knowledge redundancy are becoming more and more serious.How to obtain the interesting entertainment information efficiently and accurately has become a difficult problem in people's daily life.Recommendation system is one of the important applications in the field of data mining and machine learning.It is an important way to solve information overload.In recent years,recommendation system is more and more widely used,including news,movies,social networking sites and other fields.Personalized recommendation can build a reasonable algorithm according to the user's historical behavior data and user's attribute preference when the user has no clear demand goal,speculate the results that are most likely to meet the user's interest and feed back to the user.With the advantages of fast update and wide spread,short video can enable users to obtain a variety of information in fragmented time,which has been favored by more and more people.However,in today's era,people's life pace is fast and users' interests are changeable.Therefore,real-time recommendation for short videos is particularly important.The real-time recommendation is closely related to users' experience.Aiming at the problem of poor real-time recommendation of short video,this paper proposes a research on short video real-time recommendation method based on knowledge map.This method mainly includes two parts: the convolution representation method of knowledge map integrating time information and multiattribute features,and the improved long-term and short-term preference fusion model based on multitasking.They are introduced as follows:Firstly,Aiming at the problem of low accuracy and diversity of short video recommendation,a convolution representation method of knowledge graph integrating time information and multi-attribute features is proposed.This method first preprocesses the video through the attention mechanism,which can efficiently extract the main attribute features of the video,and then embeds the user interest by using the knowledge map.In addition to the rich node information and the association information between nodes,this paper adds the user browsing time as the weight of the relationship to measure the user's preference for the video.This method comprehensively considers the long-term and short-term interests of users,so that the recommendation results include not only the long-term interests of users themselves,but also the recent interest drift,so as to ensure the diversity and accuracy of recommendation results.Secondly,in order to solve the problem of poor real-time performance of short video recommendation,an improved long-term and short-term preference fusion model based on multitasking is proposed in this paper.Because the LSTM model has excellent performance on feature accumulation and transmission,and the changes of users' interests also have the same characteristics,this paper uses the improved LSTM model to fuse long-term and short-term interests.Take the long-term interest as the initial input of the LSTM model.When the user's short-term interest changes beyond the threshold,it indicates that the user's interest is far from the current recommendation,so the short-term interest at this time is expressed as the next input of the LSTM model.The time parameter t is introduced to represent the effective time of each recommendation,and the update gate is introduced into LSTM to illustrate the degree of interest attenuation.This model also sets up two training tasks.Based on the improved LSTM model,the training parameters of the two tasks are shared,which improves the real-time performance on the basis of ensuring the accuracy of recommendation.In order to test the effectiveness of the proposed method in this paper,comparative experiments are carried out on icme-2019 short video data set,and compared with several mainstream short video recommendation algorithms through accuracy,recall and MRR.The experimental results show that the proposed method improves the accuracy,recall and recommendation speed compared with the mainstream dynamic recommendation methods such as FM and LSTM,and improves the real-time performance of recommendation while ensuring the accuracy of recommendation.