Research And Implementation Of Video Recommendation Algorithm Based On Latent Group And Multi-modal

With the rapid development of mobile Internet technology,people have the need to watch videos every day.However,in the face of a large number of videos,how to choose the videos that the audience expects to watch has become a challenge.In recent years,there are more and more researches on video recommendation.Traditional recommendation algorithms mainly include Col-laborative Filtering(CF)and Content-based Filtering(CB).However,these two types of algorithms have problems of cold start and feature extraction.This paper will attempt to improve the accuracy of video recommendation system from two novel perspectives.This paper proposes a latent group recommendation algorithm for video.Danmaku comment function is a novel way to watch videos,which has become popular all over the world in recent years and has been provided by mainstream video websites in China.As a form of social video,users can post their thoughts on the screen while watching,and they can even communicate.The text information indirectly reflects the plot of the video.This algorithm extracts the topic distribution of video based on Danmaku comments,and assumes that users with the same preferences will gather in the same type of video because of similar topics,forming a latent group.Using the historical data of the same group of users to calculate the recommendation sequence can alleviate the problem of data sparsity to some extent.This paper proposes a multi-modal video recommendation algorithm.Movie videos usually contain rich text information such as the country of production,and at the same time,the same type of movie will contain similar sound and picture characteristics.This algorithm extracts rich text,audio and image feature information of the video respectively.Considering that online videos have the characteristics of fast update and large number,this paper sets a feature dimensionality reduction algorithm so that low-dimensional feature vectors can still carry most of the feature information,which can reduce the complexity of subsequent calculations.By matching the feature vector of the candidate video and the user-feature preference matrix in turn,the recommendation sequence is generated,and the cold start problem of the new video can be solved.By comparing with the traditional recommendation algorithm and the hybrid recommendation model,it is shown that the two video recommendation algorithms proposed in this paper can achieve good accuracy and stability in the corresponding scenes.