Research And Application Of Video Information Recommendation Model Based On Transfer Learning

Internet technology accelerates social development,and multimedia information provides people with visual enjoyment,among which video recommendation is an important part of multimedia network video service.In order to improve the recommendation quality weakened by sparse data,combining transfer learning with video recommendation has become one of the hot spots in current research solutions.Therefore,this paper studies the optimization of video recommendation model and its application,which designs and implements a video recommendation system.Main research work and achievements include:1)A video automatic encoder model based on migration hole convolution(TDCAE)is proposed.Through the classification of video sequence features,such as video category division,video numerical division and video text division,the principal component analysis technology PCA and the proposed Co-video2Vec method are used to reduce the dimension of the above divisions,so as to realize the vectorization of video feature sequence and improve data sparseness.When users browse videos in web pages or APP applications,they will generate corresponding user interactive clicking behavior sequences.Combined with the idea of cross-domain migration of source domain data to target domain,this interactive data pre-training model is used to extract potential features of users and migrate to target domain to solve data sparsity and better predict user preferences.An improved automatic encoder is proposed,and hole convolution is introduced to enhance the network receptive field and reduce the computational complexity.In the experiment,user interaction sequences with different lengths and different levels of automatic encoders are selected to explore the accuracy of the model in predicting user preferences.The experimental results on Cold Rec data sets show that the TDCAE model has better video recommendation performance than other models such as Deep FM,DIN and UPCC.2)An optimization scheme of migration recommendation based on cross-attention mechanism is proposed.By studying the deep neural network and attention mechanism,the key scoring attribute characteristics of data in the source domain are extracted,and the effective parameters of the source domain are efficiently migrated by using the cross-attention unit block without updating all the parameter weights of the model.It not only alleviates the problem of sparse scoring and cold start,but also gives higher weight to the video data that users focus on and pushes video services that focus more on users’ preferences.The experimental results show that the optimization scheme proposed in this paper has better performance.3)A video recommendation system based on transfer learning is developed by using the framework of Spring Boot+Vue.JS+MyBatis+FFMPEG.This system includes user login registration,user information management,video information management,user feedback management,video recommendation management and other functional modules.The test shows that the system designed in this paper has achieved the expected function,and its expansibility and stability are good.