Research And Implementation Of Danmaku Video Recommendation Algorithm Incorporating Multiple Features

With the rapid development of mobile Internet,people also gradually change from information access difficulties to information overload,in order to facilitate people’s efficient access to useful information,recommendation algorithms came into being.Video recommendation system has been widely used as the standard configuration of video platform,but the following problems still exist:(1)video features are not comprehensively extracted,such as comments,titles,and profiles,resulting in a lack of recommendation effection;(2)Danmaku text information is not fully utilized and integrated into video recommendation;(3)insufficient calculation of association between video content and users’ interests and preferences content,which affects the recommendation effection.In this thesis,we propose a Danmaku video recommendation algorithm integrating multiple features to address the problems in video recommendation,and the main research work is as follows:First,in order to extract the textual information which is strongly related to the video content from the video interface,this thesis uses the label information below the video homepage to get the topic distribution vector by LDA topic model training,and then uses Doc2 vec for the title and description of the video(referred to as video content profile)to get the sentence vector of the video content profile,and extracts the textual information to the video interface by these two methods.Secondly,to classify the emotion of Danmaku,a Bert＿Bi LSTM＿CNN-based approach is proposed to classify the emotion of Danmaku comments.The Bert pretrained language model with recurrent convolutional neural network Bi LSTM＿CNN is used to fully extract the Danmaku text information and sentence semantic features,and the emotion classification of the Danmaku text is derived by Softmax function,and the classification accuracy reaches 84.6%,which is a significant improvement compared with the traditional model,and the Danmaku emotion vector of each video is derived by this method.Finally,based on the emotion vector,content profile and tag topic vector,the Pearson correlation coefficient is used to calculate their respective similarities,and their weighted fusion is used to obtain the comprehensive similarity between videos,after which the decay weight is calculated by using hierarchical analysis on the user’s historical playback records,and the product of similarity and decay weight is multiplied to obtain the user’s comprehensive preference for the video,and the viewer popularity of the video is obtained by the playback index of the video homepage,which is multiplied with the comprehensive preference to obtain the recommendation value.The experimental results show that compared with the Danmaku video recommendation algorithm(DRCFT)fusing collaborative filtering and topic model and the collaborative filtering algorithm embedding LDA topic model(ULR-item CF),the accuracy of this paper’s algorithm is 12.1% higher on average,and the improvement effect is obvious.