Research On An Improved Short Video Recommendation Algorithm

This thesis focuses on an improved short video recommendation algorithm by collecting and analyzing users’ historical browsing information on short videos,targeting the interest bias generated by user behavior time on the one hand,and the impact of short video heat on the coverage of the recommendation algorithm on the other hand.Collaborative filtering is one of the most common and classic algorithms in recommendation algorithms,whose idea is to recommend targets based on the closest users or similar objects,and is often used in various recommendation systems.Some short video APPs use the traditional collaborative filtering algorithm to recommend short videos to users,but the traditional algorithm over-considers the influence of popular short videos on ratings and ignores the contribution for cold short videos on user interest attribute indexes,while taking into account the law of forgetting in human memory and borrows the Ebbinghaus forgetting curve to adjust the time decay function to track the deviation of user interest.The similarity calculation takes into account information such as the time of user behavior and the labeling characteristics of the short video itself to eliminate the cold start problem when new users or new products arrive as much as possible,so as to better ensure the quality of recommendation results and maximize the reflection of the real needs of different users to achieve personalized recommendation services.Because the traditional algorithm excessively considers the impact of popular items on the score,and ignores the contribution of popular items to the measurement of user interest characteristics,and does not consider the dynamic change of user interest.In this regard,this paper proposes a new similarity improvement algorithm.The improved collaborative filtering algorithm weights the penalty factor and time decay function of popular short videos,optimizes the user similarity calculation method,and forms a new similarity measurement model.The experimental scheme is designed using the Kuai Rec short video dataset produced by Kwai Company and the University of Science and Technology of China.An example is introduced to calculate the scoring matrix.The improved algorithm is verified by comparing with the traditional recommendation algorithm.The experimental results show that compared with the traditional recommendation algorithm,the proposed recommendation average absolute error(MAE)is significantly reduced and the recommendation quality is significantly improved.