Research On Video Recommendation Algorithm Based On Deep Learning And Data Dimension Reduction

Nowadays is an era of explosive growth of information.Video exists in people's daily life as an information carrier.Video recommendation systems can assist users quickly find suitable videos from a huge quantity of video resources,but due to some users have too few comments during video viewing,video websites cannot get user preferences based on their comments in time.Therefore,how to distinguish the true and false preferences of such users and to recommend videos with similar characteristics as their viewing history become the key topics of the recommendation system.In addition,traditional video recommendation algorithms need pay a huge storage and calculation costs,facing complex and high-dimensional data information.Therefore,how to reduce the dimension of high-dimensional video data becomes a research hots pot on the premise of maintaining the accuracy of video recommendation and meeting the personalized recommendation of users.The main research contents of this paper are as follows:(1)A high-efficiency barrage video recommendation model based on convolutional recurrent neural networks that aims to study users' true and false preferences.In order to recommend videos that meet user needs and preferences to users in a timely and accurate manner,a barrage video recommendation model based on convolutional recurrent neural network is proposed.The barrage is used to locate the user's favorite video clips in the video.The unsupervised clustering method extracts key frames from video clips,processes the key frames into effective static images,and uses them as input to the convolutional recurrent neural network model to extract important human behavior characteristics and find similar video clips.Recommend videos to users.The experimental data uses the data crawled by the Bilibili website,and the results show that the bullet screen-based convolutional recurrent neural network recommendation method mentioned in this paper can increase choosability of the user's and is compared with the traditional recommendation method in terms of recommendation accuracy.There is a big improvement.(2)Research on the two-way clustering video recommendation algorithm based on GM-LLE dimensionality reduction to enhance the accuracy of video recommendation and reduce the cost of recommendation.In order to reduce the recommendation cost,from the perspective of data dimensionality reduction,a collaborative filtering video recommendation algorithm based on GM-LLE dimensionality reduction is proposed.First,the traditional local linear embedding algorithm is improved.Although the Euclidean distance measurement in the traditional local linear embedding algorithm is simple,However,the high-dimensional space has strong complexity,and the use of Euclidean distance to measure similarity will become unreliable.So this algorithm starts with improving the distance measurement method and replaces Euclidean distance with generalized Mahalanobis distance.This method not only improves the reliability of similarity measurement,but also solves the problem that Mahalanobis distance does not exist when the matrix is singular.A two-way clustering video recommendation method that combines the characteristics of user clustering and item clustering is proposed.The experimental data uses benchmark data sets.The results show that the proposed method reduces the time complexity to a certain extent and improves the accuracy of recommendation.